[PyTorch] Allow CP P2P transport group overrides

[xiaoyao0115]

Summary

• keep the logical CP process group for CP size, rank, and peer selection
• allow CP attention P2P operations to use a separate transport process group
• route both batched and direct isend/irecv through that transport group
• store overrides weakly so registration does not extend process-group lifetimes

Motivation

Dynamic CP CUDA graphs capture multiple logical CP sizes. Initializing one NCCL communicator per logical CP group adds persistent communicator memory. The companion Megatron Core change maps those logical groups to one pre-warmed parent communicator during graph warmup and capture.

The default path is unchanged when no override is registered.

This change was split from #3353 so the transport API can be reviewed independently from CUDA Graph slot-memory sharing.

Validation

• Black 24.4.2
• Python AST validation
• transport-selection and weak-lifetime smoke test
• upstream Core and all PyTorch checks pass
• JAX and the aggregate All job fail while compiling unchanged common/activation/gelu.cu: fatbinary cannot open gelu.compute_75.ptx; this PR changes PyTorch files only

Companion change

• [dev] Reuse parent communicators for dynamic context parallelism Megatron-LM#6801
• the MCore PR is stacked on [dev] Share CUDA graph slots across dynamic CP variants Megatron-LM#6479