Fix GTP prefetch parameter readiness

[JF-D]

• I, the PR author, have personally reviewed every line of this PR.

Summary

• attach a GTP-only readiness handle to each DDP parameter-gather bucket used by GTP weights
• finish the required DDP gather before eager GTP weight reads
• record capture-local readiness dependencies in gtp_cuda_graphs.py and finish them before replaying each local forward CUDA graph
• keep the existing DDP pre-forward gather policy shared by both paths

Root cause

GTP can prefetch the next module's local weight shard before that module executes. With overlapped DDP parameter gathering, the ordinary module pre-forward hook therefore runs too late: GTP may read the distributed-optimizer parameter buffer while its bucket all-gather is still in flight. CUDA graph capture also suppresses the DDP hook, so the dependency must be retained and replayed explicitly.

Each forward graph capture now creates its own GTPCaptureCommState. The transient active state is used only while capturing; the yielded state is stored on that graph's _CudaGraphRunner, preventing handle sets from aliasing across graph instances.

[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.

[JF-D]

@claude strict-review

[JF-D]

/ok to test 0f18dfe