Makes checkpointing and evaluation bookkeeping run asynchronously

[finbarrtimbers]

Note

Extracts checkpoint saving into maybe_save_checkpoint and runs checkpointing and evaluation asynchronously after each step; updates docs with code style rules.

• Training pipeline:
  • Extracts checkpoint logic into maybe_save_checkpoint (with Timer) and replaces inline checkpointing.
  • Runs maybe_save_checkpoint and maybe_evaluate via thread pool, waits for both, then triggers weight sync.
  • Minor reordering: weight sync trigger moved to occur after async tasks complete.
• Docs:
  • Add code style guidelines in CLAUDE.md (require type annotations and Google-style docstrings).

^{Written by Cursor Bugbot for commit 74ff4e5. This will update automatically on new commits. Configure here.}

[mnoukhov]

lgtm