test(infer): swap se_e2_a fixture to deeppot_sea (type_one_side=True) (#5445)

## Summary

- Repoint the `se_e2_a` test fixture in
`source/tests/infer/deeppot-testcase.yaml` from the legacy
`deeppot.yaml` (`type_one_side=False`) at `deeppot_sea.yaml`
(`type_one_side=True`).
- Regenerate reference values for the 3 result blocks (PBC, NoPBC,
big-box) plus descriptor and fit_ll for block 1 using the same
`convert_backend → tmp.pth → DeepPot.eval(...)` pipeline that
`TestDeepPot` uses at runtime.
- Drop the now-unneeded `if key in ("se_e2_a", "se_e2_r") and extension
in (".pte", ".pt2"): SKIP` clause; only `se_e2_r` still skips for
pt_expt extensions.
- Remove the obsolete `source/tests/infer/deeppot.yaml` (no remaining
references; `deeppot.pbtxt`, used directly by C++ tests, is unrelated
and untouched).

## Why

The legacy `deeppot.yaml` used `type_one_side=False`, which hits a
data-dependent symbolic-int guard in `make_fx` when pt_expt tries to
export the `ndim=2` NetworkCollection. As a result,
`TestDeepPot_se_e2_a_.pte` and `TestDeepPot_se_e2_a_.pt2` skipped
entirely, leaving the plain `se_e2_a → .pte / .pt2` slice of `dp
convert-backend` uncovered. Only `fparam_aparam_.pte/.pt2` exercised the
pt_expt export path, and that case carries extra fparam/aparam
complexity.

After the swap, all four extensions (`.pb`, `.pth`, `.pte`, `.pt2`) run
for `se_e2_a`. `dp convert-backend deeppot_sea.yaml → .{pb,pth,pte,pt2}`
is now exercised end-to-end with energy / force / virial assertions,
plus descriptor (.pth, .pb) and fit_ll (.pth) pinning.

## Test plan

- [x] `python -m pytest source/tests/infer/test_models.py` → `71 passed,
51 skipped` (was `55 passed, 67 skipped`).
- [x] `python -m pytest -v -k TestDeepPot_se_e2_a_` → all 40 cases
either PASS or SKIP for documented reasons (test_descriptor /
test_fitting_last_layer skip on .pte/.pt2 per the existing global skip
in `test_models.py`).
- [x] `test_model_script_def` passes on all 4 extensions, confirming the
case yaml's `model_def_script` matches what's serialized in
`deeppot_sea.yaml` (and so survives `convert_backend` round-trips).

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Tests**
* Updated a test fixture to use a variant-specific model definition and
simplified model configuration.
* Regenerated numeric result data for energies, forces, virials,
descriptors, and fit outputs.
* Added per-key precision/tolerance control and applied it across
comparisons.
* Adjusted test skip/exclusion behavior for certain model export
formats.

<!-- review_stack_entry_start -->

[![Review Change
Stack](https://storage.googleapis.com/coderabbit_public_assets/review-stack-in-coderabbit-ui.svg)](https://app.coderabbit.ai/change-stack/deepmodeling/deepmd-kit/pull/5445)

<!-- review_stack_entry_end -->
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Han Wang <wang_han@iapcm.ac.cn>

Author: wanghan-iapcm