perf(fsdp): skip deterministic NaN fill in moves; fix reference worker PYTHONPATH

[Rockdu]

Stack 5/9 of follow-ups to #263: main ← #265 ← #266 ← #267 ← #268 ← this ← #271 ← #272 ← #273 ← #274.

What

• The GPU reference-model test puts the repository root on its worker's PYTHONPATH, so it also runs outside CI (the worker imports the CPU test's harness from tests). refactor(lora): one plain PEFT adapter directory per component #274 keeps the caller's PYTHONPATH behind it instead of replacing it.
• Model and optimizer moves turn off deterministic mode's NaN fill of uninitialized memory while they run: every tensor they allocate is overwritten by its copy, so the fill only cost time. The saving has not been measured on its own.

Not in this PR

• sleep_wake.py still reads the singular FSDPState._fsdp_param_group. The plural _fsdp_param_groups exists only from torch 2.12, and the repo pins torch 2.11 (the Wan2.2 param-dtype patch refuses other versions), so the switch belongs to the torch upgrade.

Files

• miles/backends/fsdp_utils/sleep_wake.py: _skip_uninitialized_memory_fill.
• tests/fast-gpu/backends/fsdp_utils/test_reference_models.py: worker PYTHONPATH.

Checklist

• pre-commit run --all-files passes — not run locally
• Added/updated tests for new behaviour — no new test; e2e standards verify the deterministic runs are unchanged
• pytest -x is green — GPU tests run in CI only
• If launch flags changed, python3 train.py --help still parses — n/a
• If a public flag was added, it appears in the CLI reference docs — n/a
• If an example was added, it has a real walkthrough — n/a