perf(events): SSE dirty-flag + tear-proof job serialization

[thcp]

Summary

• PR-3A of Phase 3 (Performance). Adds Job.version, bumped by _set() on every field write. Internal-only (excluded from to_state() and to_record(), same as cancel_requested).
• job_events's SSE loop now compares job.version to the last-sent version instead of re-serializing + string-diffing to_state() on every 0.2s tick — idle connections do one int compare per tick instead of a full to_state() + json.dumps(). At the 200-connection cap that removes ~1,000 serializations/s.
• Closes Non-atomic multi-field job updates visible to SSE #285 for real: if job.version changes while to_state() is mid-call (a concurrent _set() from the pipeline thread), the snapshot may mix pre- and post-write fields — a torn read. The loop now detects that (version read before vs. after serializing) and discards the snapshot instead of yielding it, retrying immediately rather than sleeping.
• Already-terminal jobs (done/error/cancelled) close the stream right after the initial snapshot instead of idling on int-compares.

Closes #289

Test plan

• New tests/test_events_stream.py drives the SSE generator directly: initial snapshot always sent, no duplicate frames while state is static, a frame is emitted on _set() and the stream closes after a terminal transition, a simulated torn read (monkeypatched to_state() bumping version mid-call) is discarded and re-serialized consistently, an already-terminal job closes after one frame.
• Full suite: 188 passed, 12 skipped
• ruff check / ruff format --check clean