feat: add replay preflight verificationFeat/replay preflight

[pearlq12345]

Summary

Extends the existing preflight verifier from inference-only to inference + replay. Adds replay-specific argv and parameter validation that runs before start_replay() / run_replay() spawn the subprocess.

What it validates (replay path)

• wrapper action must be replay
• required replay dataset args present in argv (repo_id, root, episode, fps)
• episode >= 0
• 1 <= fps <= 120
• at least one follower arm in manifest
• bimanual replay still requires left/right consistent followers

Why

Keeps the verifier aligned with the existing "host-visible checks before subprocess creation" direction, and extends the infer preflight pattern to replay without touching train.

Self-review (against PR #62 feedback)

• Tightened _MAX_REPLAY_FPS from 240 → 120 (more realistic bound for real replay)
• Naming mirrors the infer convention (_verify_replay_preflight, _validate_replay_*)
• Validation set focuses on web/service entrypoint; no CLI-only edge cases

Validation

• python -m compileall and git diff --check pass
• New unit tests for valid replay, bad replay argv, and invalid limits
• New lifecycle integration test confirms preflight rejects bad params before subprocess start (subprocess never spawned)

Notes

• This PR only extends replay.
• Intentionally not touching train to avoid conflict with ongoing work there.