feat: enable MACE pt_expt torch.compile training

[njzjz]

Summary:

• add MACE compatibility for DeePMD-kit pt_expt lazy torch.compile tracing
• use the exportable dense-edge/raw-MACE path during make_fx tracing while sharing the original trainable parameters
• document training.enable_compile usage and default-config steady-state benchmark results

Benchmark:

• command path: srun --gres=gpu:1 dp --pt-expt train input.json
• input: water example data with sel=auto, hidden_irreps="128x0e + 128x1o", batch_size=1, disp_freq=100; 2000 steps were used only to sample stable per-step timings
• eager steady-state after step 200: 0.2030 s/step
• torch.compile steady-state after step 200: 0.1495 s/step
• steady-state speedup: about 1.36x
• first compiled step included a one-time Inductor trace/compile cost of 97.17 s, amortized over long training runs

Tests:

• ruff check deepmd_gnn/mace.py
• pytest -q tests/test_pt_expt.py
• git diff --check
• pre-commit hooks on commit, including mdformat

[codecov]

Codecov Report

❌ Patch coverage is 22.05882% with 53 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.97%. Comparing base (33be5a2) to head (ff22c86).

Files with missing lines	Patch %	Lines
deepmd_gnn/mace.py	22.05%	53 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #134      +/-   ##
==========================================
- Coverage   83.85%   80.97%   -2.88%     
==========================================
  Files          12       12              
  Lines        1406     1472      +66     
==========================================
+ Hits         1179     1192      +13     
- Misses        227      280      +53     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.