refactor(distributed): make handle_fn and launch_nowait share the call halves

[haonan3]

Summary

Extract _launch_half (dispatch → localize → execute) and _collect_half (ray.get → rebind → collect) on Handle; handle_fn, launch_nowait, and PendingHandleCall.result now sequence the same two methods instead of carrying keep-in-parity copies — the "keep the halves in parity" comment convention (#287's own listed follow-up) is deleted. No behavior change: the grad-context preamble/postamble stays in handle_fn; launch_nowait stays grad_mode=False/call_id=None.

Split out of #295 per review: this is a distributed-RPC-layer change with its own reviewers and regression surface, independent of the rollout refactor that consumes the seam.

Related Issue

N/A (follow-up to #287).

Test Plan

• ruff format --check + ruff check over unirl/distributed/group/handle.py — clean; python3 -m compileall — clean.
• Pure extraction verified by line-diff: both halves are byte-identical to the previous inline bodies; call sites are thin sequencing.
• GPU: Not run; reason: CPU-only dev box. Any sync + async smoke exercises both entries (handle_fn on every @distributed call; launch_nowait on the async trainers); the refactor(rollout): dedup async driver plumbing; extract ToolAgentHarness #295 validation trio covers it.

Compatibility / Risk

No public surface change; _execute_all/_rebind_tree/_method_configs untouched. Riskiest spot: none identified — the halves are verbatim relocations.

Reviewer Notes

AI-assisted (Claude Code). Suggested land order: this → #294 → #295 (all touch adjacent layers; each rebases mechanically).

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.

[haonan3]

Folding back into #295 per team decision — the handle-halves commit stays cleanly separated as that PR's first commit; reviewers can review it per-commit there.