fix(ep): drain CXI FIFO rings before quiet

[fergusfinn]

Summary

• Post and wait QUIET across every D2H ring on the CXI path so sibling-ring work cannot remain queued behind a completed quiet.
• Keep BARRIER as a single rank-level rendezvous after the quiet fence.
• Cache CXI transport mode per proxy and adjust the 8-rank combine config to satisfy the forwarder invariant with NUM_MAX_NVL_PEERS=4.

Testing

• git diff --check
• python3 -m py_compile ep/bench/buffer.py
• 2-node DeepSeek-V4-Flash vLLM serve, CXI transport, NUM_MAX_NVL_PEERS=4
• vLLM benchmark: ISL=1024 OSL=1024 CONCURRENCY=4096 NUM_PROMPTS=4096 REQUEST_RATE=inf, completed 4096/4096, failed 0, output throughput 6132.09 tok/s, total throughput 12288.13 tok/s

[Copilot]

Pull request overview

This PR aims to make nvshmemi_ibgda_quiet correctly fence/drain CXI D2H work that is sharded across multiple rings by posting (and waiting on) QUIET commands across all relevant D2H channels, instead of one ring per proxy thread.

Changes:

• Adds a CXI-specific stride/limit so QUIET can be posted on every D2H ring (vs. one per proxy).
• Tracks the exact D2H channel index for each posted QUIET so the wait phase targets the same rings that were posted.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.