Fix/fly opt gpu llvm translations

[Peter9606]

Summary

• Add a fly-opt regression smoke test that exercises the real gpu-module-to-binary path on a minimal gpu.module.
• Register GPU-to-LLVMIR translation interfaces in fly-opt, matching the requirements of MLIR GPU binary serialization.
• Link MLIRToLLVMIRTranslationRegistration so the translation registration symbols are available.

Motivation

fly-opt currently registers MLIR dialects, extensions, and passes, but not the GPU-to-LLVMIR translation interfaces required by gpu-module-to-binary.

As a result, running binary codegen through fly-opt can fail before target serialization with:

missing LLVMTranslationDialectInterface registration for dialect for op: gpu.module

Test Plan

FLYDSL_FLY_OPT=/tmp/flydsl-iluvatar-runtime-smoke/bin/fly-opt \
  python3 -m pytest tests/unit/test_fly_opt_tooling.py -q --confcutdir=tests/unit
Result:

1 passed
When no fly-opt binary is available, the test skips unless FLYDSL_FLY_OPT or FLY_OPT points to a built binary.

[Peter9606]

Could @coderfeli please take a look at this PR? This fixes a generic fly-opt registration gap exposed by gpu-module-to-binary, and I would appreciate a review on whether this is the right place to register the GPU-to-LLVMIR translations.