refactor(lora): one plain PEFT adapter directory per component

[Rockdu]

Stack 9/9 of follow-ups to #263: main ← #265 ← #266 ← #267 ← #268 ← #269 ← #271 ← #272 ← #273 ← this.

What

• --lora-adapter-path, --ref-lora-adapter-path and --teacher-lora-adapter-path take one plain PEFT adapter directory (what save_pretrained writes) per --update-weight-target-module entry, in order, e.g. --lora-adapter-path /adapters/high_noise /adapters/low_noise for Wan2.2. Any other count is rejected in arguments.py.
• The component-named subfolder layout and its lora_adapter_subfolder helper are gone.
• test(fsdp): keep the caller's PYTHONPATH ...: the GPU reference-model test now puts the repository root in front of the caller's PYTHONPATH instead of replacing it (a follow-up to perf(fsdp): skip deterministic NaN fill in moves; fix reference worker PYTHONPATH #269, kept here so the PRs below keep their CI results).

Why

#263 read a multi-component adapter as one directory with a PEFT checkpoint under each component subfolder. No other tool writes or reads that layout: PEFT names its subfolders after adapters, and diffusers / the Wan2.2 community ship one file per expert. A list of standard PEFT directories keeps the loader on PEFT's own format.

Files

• miles/utils/arguments.py: the three flags become nargs="+" lists (args.*_lora_adapter_paths); the count checks run once the component list is parsed.
• miles/backends/fsdp_utils/model_loader.py, actor.py: each component loads its own adapter directory.
• tests: test_reference_arguments.py (count check), test_reference_models.py and the GPU reference worker (per-component directories); tests/fast-gpu/backends/fsdp_utils/test_reference_models.py (PYTHONPATH).
• docs: fsdp_model_roles.md, cli-reference.md.

Checklist

• pre-commit run passes on the changed files
• Added/updated tests for new behaviour
• pytest -x is green — test_reference_arguments.py, test_reference_models.py locally; the GPU worker runs in CI
• If launch flags changed, python3 train.py --help still parses — argparse nargs="+" checked standalone; the full parser runs in CI
• If a public flag was added, it appears in the CLI reference docs
• If an example was added, it has a real walkthrough — n/a