MFSDP v2: Support EP composability with grouped 3D expert weights

Follow-up to #5656.

#5656 describes the grouped-expert representation: a single contiguous 3D expert-weight tensor, such as `[E_local, h_in, h_out]` or `[E_local, h_out, h_in]`, rather than separate 2D `torch.Parameter`s per expert.

This issue tracks implementing EP + MFSDP v2 support for that representation.

## Motivation

The grouped contiguous representation simplifies OSS MoE model implementations and can enable optimizations that are difficult to guarantee when each expert weight is an independent parameter.

Start by reading #5656 to understand the grouped-expert representation, then examine how expert parallelism and MFSDP v2 currently interact. The issue does not name files, tests, or entry points, so the implementation scope and completion criteria need to be established before coding; done means EP supports the contiguous 3D expert-weight representation with MFSDP v2.

Written by the indexing model from the issue text.