feat(training): enable MFSDP V2 gradient clipping

[wujingyue]

Summary

• allow clip_grad > 0 for MFSDP V2
• let the dense and EP=2 V2 smoke tests use the normal clipping default (1.0)

Dependencies

• Depends on the MFSDP V2 EP integration in feat(training): support MFSDP V2 expert parallelism #5521.
• Depends on MCore EP support in feat(fsdp): support expert parallelism in MFSDP v2 adapter NVIDIA/Megatron-LM#6450.
• MCore clipping coverage is added separately in fix(optimizer): support MFSDP v2 gradient clipping NVIDIA/Megatron-LM#6489.

Validation

• MCore two-rank dense and EP=2 clipping coverage in fix(optimizer): support MFSDP v2 gradient clipping NVIDIA/Megatron-LM#6489.
• uv run pre-commit run --all-files
• Bridge two-rank smokes were not runnable in this worktree: its committed MCore submodule is intentionally pinned before #6450, and the isolated local environment failed while building fast-hadamard-transform because it does not declare Torch as a build dependency.

[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.