fix(grpo): minor issues (regex whitespace, CoT adapter reuse, shutil import)

[abrichr]

Collection of Small Issues

1. CLICK regex (trainer.py:761) doesn't handle whitespace variants: CLICK( x=0.5, y=0.5 ) won't match. Add optional whitespace: CLICK\s*\(.

2. CoT warmup (cot_warmup.py:93) creates a new API adapter per step inside the annotation loop. Should be hoisted outside the loop.

3. import shutil inside save_checkpoint() is re-imported every call (minor).

4. Port mismatch: GRPOConfig.server_url defaults to port 5001 but WAALiveConfig defaults to 5000 — document this difference or unify the default.

[abrichr]

Closing: resolved by replacing the custom GRPO trainer with a minimal TRL bridge (see PR #34). The custom parsing code (CLICK regex, CoT warmup adapter reuse, shutil import, port mismatch) is eliminated — TRL replaces the custom trainer.py where these issues existed. The port default concern may still be relevant at the evals layer but is not an openadapt-ml issue.