[gbesposito]

What does this PR do?

This PR fixes and speeds up the EP integration for MoE models

General Changes

• Device mesh: EP folded into dp_shard; expert mesh (efsdp, ep) built via _unflatten; EP + TP/PP/CP rejected for now.
• FSDP: experts FSDP-sharded with the correct gradient divide factor, fp32 master weights and explicit prefetching.
• Token dispatch and experts: a single host sync per layer and a GPU-side token permutation; grouped_mm expert backend).
• MoE model: dropless routing by default; fixed capacity formula (now includes top_k) and expert init.
• Global-batch load balancing
• EPAdamW removed: standard AdamW with param groups split by mesh, a per-mesh gradient clipper, and the standard DCP path
• MoE MFU Calculator (active params only) and a qwen_moe init group.

Breaking Changes

• expert_parallel_degree now divides dp_shard instead of multiplying the world size.
• ep_adam_w is removed; use adam_w.
• The ep_wrapped component no longer takes mixed_precision_settings.
• TransformerBlock.aux_loss is removed. MoECrossEntropyLoss takes device_mesh.
• moe_capacity_factor defaults to None (dropless).
• Use model_type: qwen_moe for init and the moe MFU calculator.

Checklist before submitting final PR

• My PR is minimal and addresses one issue in isolation
• [x ] I have merged the latest version of the target branch into this feature branch
• [ x] I have reviewed my own code w.r.t. correct implementation, missing type hints, proper documentation, etc.
• [x ] I have run a sample config for model training
• [ x] I have checked that all tests run through (python tests/tests.py)
• I have updated the internal changelog (CHANGELOG_DEV.md)