fix: export operator call instantiations

[voltjia]

Summary

• Reverts the public infini::ops::functional layer added by feat: add public C++ operator API #618.
• Generates explicit Operator<Op>::Call template instantiations into libinfiniops.so and generated extern template declarations for C++ consumers.
• Restores #include <infini/ops.h> as the public C++ entrypoint for existing operator classes and adds an external C++ smoke test for infini::ops::Add::Call.

Motivation

Closes #593

Downstream C++ consumers should not need backend kernel headers or vendor compilers just to call existing Operator<Op>::Call APIs. Explicit template instantiation keeps the existing operator class API and moves backend-dependent instantiation into libinfiniops.so, avoiding the extra functional wrapper layer.

Rebase Status

• Rebased onto latest master: 42738491b4306f247c59eba24aaffc1001885cdb (ci: stabilize iluvatar runner and test images (#625)).
• Current head: e870e3e74b196226ec80cd8ff9f843f838a70f39.

Type of Change

• N/A — feat: this does not add a new feature/operator/platform.
• fix — bug fix.
• N/A — perf: no performance-path change.
• N/A — refactor: not a behavior-neutral restructuring only.
• N/A — test: includes tests but is not test-only.
• build / ci — build system/codegen/linkage configuration.
• N/A — docs: not documentation-only.
• N/A — Breaking change: this restores the existing Operator<Op>::Call surface rather than introducing an ABI break.

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Full Platform Test Results

All accelerator runs used card 6. Commands installed the package first, then ran bare pytest with no test-path or device arguments.

Platform	Build	pytest result	Device / Notes
NVIDIA	Yes	9207 passed, 8665 skipped, 81 warnings in 342.43s	ssh nvidia, Docker --gpus "device=6", in-container CUDA_VISIBLE_DEVICES=0
MetaX	Yes	8699 passed, 7655 skipped, 81 warnings in 399.34s	ssh metax, CUDA_VISIBLE_DEVICES=6
Iluvatar	Yes	7705 passed, 8649 skipped, 81 warnings in 582.92s	ssh iluvatar, CUDA_VISIBLE_DEVICES=6
Moore	Yes	8472 passed, 7900 skipped, 99 warnings in 618.44s	ssh moore, MUSA_VISIBLE_DEVICES=6; required LD_PRELOAD=/usr/local/musa-4.3.1/lib/libomp.so because /usr/local/musa/lib/libomp.so does not export __kmpc_for_static_fini
Cambricon	Yes	5900 passed, 10070 skipped, 172 warnings in 978.34s	ssh cambricon, MLU_VISIBLE_DEVICES=6, command exit status 0
Ascend	Yes	7398 passed, 8914 skipped, 71 warnings in 632.45s	ssh ascend, ASCEND_RT_VISIBLE_DEVICES=6; pytest reached a passing summary, but the outer Docker command wrote exit status 137 after completion, so this is recorded as an environment/teardown anomaly rather than a clean command pass

Validation commands

# Rebase and push
ssh nvidia 'cd /home/huangjiacheng/infiniops-pr623-rebase && git fetch origin master && git rebase origin/master'
ssh nvidia 'cd /home/huangjiacheng/infiniops-pr623-rebase && git push --force-with-lease origin fix/export-call-instantiations'

# NVIDIA
python -m pip install .[dev] --no-build-isolation && pytest

# MetaX
pip install .[dev] --no-build-isolation && pytest

# Iluvatar
python -m pip install packaging exceptiongroup typing-extensions pygments pybind11 libclang
python -m pip install . --no-build-isolation --no-deps && pytest

# Moore
export LD_PRELOAD=/usr/local/musa-4.3.1/lib/libomp.so
pip install .[dev] --no-build-isolation && pytest

# Cambricon
pip install .[dev] --no-build-isolation && pytest

# Ascend
pip install .[dev] --no-build-isolation && pytest

Benchmark / Performance Impact

N/A — this PR changes build/codegen/linkage for C++ operator calls, not operator kernels or performance paths.

Notes for Reviewers

• Public C++ consumers should include <infini/ops.h> to get the generated extern template declarations before calling infini::ops::<Op>::Call.
• The generated instantiation sources include backend marker and implementation headers, so downstream consumers no longer need backend kernel headers for the covered Call signatures.
• Operator::Call now takes const Args&... to make the instantiation signature stable across lvalue/rvalue call sites.
• Ascend produced a complete passing pytest summary, but the outer Docker process returned 137 after pytest completed. The test result is included for review transparency instead of being marked as a clean command pass.

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz where <type> matches the PR title's Conventional Commits type.
• Each commit message follows Conventional Commits.
• Every commit is meaningful, well-formed, and independently reviewable.
• No stray merge commits from master.
• No fixup! / squash! / wip commits remain.

Scope and Design

• Changes are minimal and scoped to reverting functional plus exporting Call instantiations.
• No dead code, debug prints, or unowned TODOs were added.
• No unrelated formatting churn was introduced.
• Public API changes are intentional and covered by the external C++ smoke test.

General Code Hygiene

• Comments were added only where the behavior is non-obvious.
• Modified and added files end with trailing newlines.
• No trailing whitespace, tab/space mixing, or stray BOMs.
• Identifiers in comments and error messages use backticks where applicable.
• Comments and error messages are in English.
• Comments and error messages follow project conventions.

C++ Specific

• Code follows the Google C++ Style Guide.
• clang-format --dry-run --Werror include/infini/ops.h src/operator.h passed on ssh nvidia in the PR creation pass.
• N/A — clang-tidy was not run in this pass; this PR does not add new kernel logic.
• N/A — operator parameter order is unchanged.
• No exceptions are thrown.
• No new error or warning messages were introduced.
• N/A — no kernel files were added or renamed.
• N/A — no kernel/kernel launcher split was changed.
• N/A — no constructor initializer lists were changed.
• Namespace and spacing in touched C++ headers are formatted by clang-format.
• N/A — no new operators were added.
• No raw new/delete was introduced.

Python Specific

• ruff check scripts/generate_wrappers.py tests/test_cpp_api.py passed in the PR creation pass.
• ruff format --check scripts/generate_wrappers.py tests/test_cpp_api.py passed in the PR creation pass.
• Comments and strings follow surrounding Python conventions.
• Framework-specific pytest.skip conventions are preserved.
• Function-body spacing follows the surrounding style.
• Control-flow spacing follows the surrounding style.
• Return statement spacing follows the surrounding style.
• N/A — no docstrings were added.
• N/A — the touched generator/test code follows the existing no-type-hints style.

Testing

• Full bare pytest was run on all six supported accelerator platforms: NVIDIA, MetaX, Iluvatar, Moore, Cambricon, and Ascend.
• All platform results are listed in the table above, including the Ascend outer-process 137 teardown anomaly.
• New functionality has a regression smoke test under tests/.
• N/A — the new smoke test does not need parametrization.
• N/A — pytest.mark.auto_act_and_assert is not applicable to this external compile/link smoke test.
• N/A — dtype/device parameterization is not applicable to this external compile/link smoke test.
• N/A — no new flaky test behavior was observed in the passing pytest summaries.
• The external Add::Call smoke fails on the previous behavior with an empty backend dispatch and passes with this PR.

Build, CI, and Tooling

• pip install/wheel build was run as part of every full-platform pytest command above.
• N/A — compile_commands.json regeneration was not separately checked.
• N/A — no new backend/device was added.
• Existing CUDA-like backend mutual-exclusion logic was not changed.
• Local equivalents of clang-format.yml and ruff.yml checks passed for touched files in the PR creation pass.
• No new runtime dependency was added.

Documentation

• N/A — README/CONTRIBUTING updates are not required for this internal codegen/linkage fix.
• N/A — no new operator or public utility API was added beyond restoring <infini/ops.h> as the entrypoint.
• N/A — no breaking change is introduced.

Security and Safety

• No secrets, access tokens, internal URLs, customer data, or personal hardware identifiers were committed.
• N/A — no third-party code was added.
• No unsafe pointer arithmetic, uninitialized reads, or missing bounds checks were introduced.

[voltjia]

请 @crapromer 初审，@Ziminli 终审。