Enable GTP topology for heterogeneous MIMO training

[yashaswikarnati]

What

• Resolve dense and expert weight-shard counts against the language TP axes.
• Materialize GTP/EGTP dimensions and all required module-owned process groups in the heterogeneous grid.
• Account for GTP data lanes in bridge fan-out, mock data, token normalization, and batch sizing.
• Seed dense and expert GTP ranks explicitly without MPU-global process-group reads.
• Reuse each module's ProcessGroupCollection during optimizer construction.
• Keep encoder and projection configs non-GTP.
• Extend the 20-layer mock recipe to support the multi-node topology required by GTP8/EGTP2.

Dependency

Stacked on #6234 by @fanshiqing. Checkpoint semantics are intentionally excluded and follow in a separate PR.

Validation

• behavioral argument and process-group topology tests
• Python compile checks
• shell syntax check
• import sorting
• git diff --check
• full 100-iteration save/load validation 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.