[Pipeline] Enforce an outstanding send budget (#196727)

## Summary

Add an optional `max_outstanding_sends` limit to runtime pipeline schedules. The schedule moves existing send waits earlier to keep pending forward and backward sends within the per-rank limit.

Waits move only when dependency analysis proves the placement cannot create a cycle. Unsafe limits raise an actionable error. The default behavior remains unchanged.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/196727
Approved by: https://github.com/sanketpurandare
ghstack dependencies: #196640, #196724, #196725, #196963, #196726