feat: bound pipeline work items with a per-item timeout

[lollllcat]

Summary

• Add WORKITEM_TIMEOUT_SEC (default 120min, override via CRISP_WORKITEM_TIMEOUT_SEC) to crisp/pipeline.py: a work item running longer than this gets its worker process force-killed (terminate(), escalating to kill() if needed) and its Config marked failed, instead of blocking the pipeline stage forever.
• pipelineWorkerReal's wait loop now bounds mp.connection.wait(...) by the soonest deadline and reaps expired workers each time it times out. Only applies to non-serialize (multiprocess) phases.
• Extracted the existing out-of-order completion flush logic into a shared _flushReadyItems() helper.
• Added tests for the timeout/kill-escalation/reaping logic, plus an end-to-end test confirming a hung work item gets reaped instead of hanging the stage.

Test plan

• pytest full suite passes (647 passed)
• bazel test //... passes (33/33 targets)