Improve compute scheduler throughput and reduce lock contention

- Parallelize per-runner SubmitActions via WorkerThreadPool when
  multiple runners are active
- Make remote submissions non-blocking: dispatch to pool and return
  immediately, marking actions as Submitting so the scheduler won't
  re-pick them; failures transition to Failed for auto-retry
- Downgrade m_ActionMapLock from exclusive to shared in
  SchedulePendingActions and move sort/resize outside the lock
- Release m_RunnersLock before submitting by snapshotting runners
  as Ref copies, avoiding holding the lock during HTTP I/O
- Add per-runner submit timing warning (>500ms) with host context
- Add GetDisplayName() virtual to FunctionRunner for diagnostics

0 files changed, 0 insertions, 0 deletions