Pass process groups to remaining GTP-backed layers

[yashaswikarnati]

What

• Pass the owning ProcessGroupCollection to the Mamba input/output projections.
• Pass it through the shared-expert MLP.
• Pass it to the hybrid model embedding and output projection.

Position embeddings and other small parameters remain replicated.

Why

#6234 adds the GTP-capable layer APIs, but these MIMO-used construction sites still supplied only TP. This completes the explicit process-group path for their dense GTP-rematerialized weights without adding MPU-global reads.

Dependency

Stacked on #6234 by @fanshiqing.

Validation

• Python compile checks for changed files
• import sorting
• git diff --check
• full GPU E2E is tracked by the integration PR

[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.