Fix KernelCh completion ordering in profiler proxy

[usernamehaha2022]

Description

This PR fixes a race in the NCCL profiler proxy where a KernelCh completion event could be processed before the corresponding start event was delivered.

The proxy now only stops/drains a KernelCh profiler sub-operation after StartKernelChEvent has been posted.

In stress testing with many small Ring+LL collectives, inspector logs showed missing coll_sn records across ranks, indicating dropped profiler records caused by this ordering race. Example gaps included:

Related Issues

N/A

Changes & Impact

1. Adds a guard so ncclProfilerStopKernelChEvent only runs after sub->posted == sub->nsteps.
2. Reuses the computed profiler ring index instead of recomputing it.
3. No API or ABI changes.

Performance Impact

No expected impact on normal NCCL collective performance.
The change is limited to the profiler proxy path and only adds one CPU-side condition when profiler/inspector is enabled.

[xiaofanl-nvidia]

++ @armratner @rishdas to take a look

[armratner]

This no longer will be relevant, as we decouple from the proxy in 2.31, and we covered this issue.