[dev] Reuse parent communicators for dynamic context parallelism

[xiaoyao0115]

Summary

• keep logical dynamic-CP groups for rank/topology, but reuse the existing parent DP×CP communicator for Transformer Engine P2P attention
• route dynamic-CP router subgroup reductions through fixed, disjoint lanes on the same parent communicator
• avoid eagerly initializing a separate NCCL communicator for every logical CP-size group when the parent transport is available

The logical groups still determine CP size, rank, and peers. Only the communication transport is reused, so non-P2P and local implementations keep their existing behavior.

Implementation

MCore registers each logical CP group with its parent DP×CP group before TE attention. TE selects peers from the logical group but issues P2P on the parent ProcessGroup. Router reductions use a parent-sized fixed layout so different dynamic subgroups do not overlap.

This works independently of CUDA Graph. CUDA Graph slot/memory sharing is the follow-up in #6479.

Dependencies and merge order

1. [PyTorch] Allow CP P2P transport group overrides TransformerEngine#3420
2. this PR
3. [dev] Share CUDA graph slots across dynamic CP variants #6479

This PR targets dev; #6479 is stacked on pull-request/6801.

Validation

• Black and isort checks
• 11 targeted communicator/router unit tests
• 4-GPU dynamic-subgroup reduction tests
• combined 4-GPU CUDA Graph probe: 50 replays with mixed CP2/CP1/CP1

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[xiaoyao0115]

/ok to test d03753d

[xiaoyao0115]

/ok to test cd8e9b2

[xiaoyao0115]

/ok to test d7cb2de