Add remote L3 worker

[puddingfjz]

Summary

• Add a Remote L3 worker design for scheduling remote Worker(level=3) instances as L4 NEXT_LEVEL children while preserving the existing Orchestrator/Scheduler DAG model
• Introduce a WorkerEndpoint boundary so local fork/shm children and remote L3 sessions share the same run/control/shutdown contract
• Define fork-safe remote daemon/session startup, including runner prestart before HELLO READY and before any transport or health threads handle task traffic
• Specify remote callable routing with separate outer remote-orch and inner L3 cid namespaces, transactional register/commit/abort, unregister tombstones, and unified
  register(callable, workers=...) API semantics
• Define remote TaskArgs wire representation with canonical frame encoding, remote tensor sidecars, tag consumption before dispatch, explicit remote buffer handles, and first-cut
  rejection of null remote OUTPUT tensors
• Define failure semantics for explicit success/failure completion, endpoint failure, downstream dependency poisoning, timeout handling, and cleanup
• Specify HCOMM-backed steady-state control/data communication after bootstrap, including HCOMM RPC command lanes, HCOMM data adapter profiles for A2 RoCE/A3 HCCS/A5 UB, and
  independent health signaling
• Keep Remote CommDomain and possible CommContext ABI changes as later work while reserving domain-related controls for future extension
• Provide a staged implementation plan and test matrix covering local adapter compatibility, frame validation, remote simulation, callable registration, remote buffers, health, and
  hardware-gated HCOMM profiles

Testing

• Docs only; no tests run
• git diff --check passed

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!