Fix layer-wise (Muon) distributed optimizer + fp8-param

[fanshiqing]

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

What does this PR do?

Support Muon (LayerWiseDistributedOptimizer) with mxfp8 fp8-param-gather
fp8-param-gather relies on the DistributedOptimizer param-buffer machinery, which
LayerWiseDistributedOptimizer (Muon) does not have, so the optimizer's weight
update never reached the forward and loss stayed flat. Wire the missing pieces so
Muon trains with --fp8-param-gather + --reuse-grad-buf-for-mxfp8-param-ag at both
DP=1 and DP>1:

• optimizer.py: the reuse_grad_buf (master -> fp8 param buffer) copy only applies
  to DistributedOptimizer; gate it on _copy_main_params_to_param_buffer so
  non-DistOpt optimizers (Muon's Float16 base opts) fall through to
  master -> model params. The forward then reads the update.

• param_and_grad_buffer.py: in the LayerWise variable-size param all-gather, stage
  the high-precision fp32 master (param.main_param, cast to bf16) for MXFP8 params
  instead of flattening the MXFP8 wrapper (view(-1) is unsupported); the receive
  side re-quantizes via param.data.copy_. Also skip _post_param_sync for buckets
  with no param_data buffer (the LayerWise path already updated params in place).

[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.

[github-actions]

This PR has been automatically converted to draft because all PRs must start as drafts.

When you are ready for review, click Ready for Review to begin the review process. This will:

1. Add the oncall reviewer (optional reviewer)
2. Add required review teams based on your changes

See the contribution guide for more details.

[fanshiqing]

/ok to test b9ebad6

[fanshiqing]

/ok to test eb367a4

[fanshiqing]

Close this MR as this fix (fanshiqing@1b0121d) has already been merged with #4967 together.