fix(tui): job queue hardening — partial failure rollback + batch health checks

[seilk]

Post-merge hardening for Job Queue (feature/job-queue-v1)

From Capybara review R2 recommended items:

R1: Partial Failure Rollback in executeJobRemote

• One-to-one mode: if GPU N fails mid-launch, all previously started tmux sessions are now killed (rollback)
• Prevents orphaned sessions consuming GPU resources on partial failure

R3: Batch Health Checks in watchRunningJobs

• Replaced per-job Python process spawning with batchCheckJobsAlive()
• All running jobs checked in a single Python invocation per watchdog cycle
• Reduces overhead from N cold Python starts to 1 (significant with 10+ running jobs)

Testing

• 34/34 existing tests passing
• No behavioral change for single-job scenarios (checkJobAlive preserved as thin wrapper)

Ref: CAPYBARA_REVIEW_R2.md (in raven worktree, now removed)