Support uneven communication for MFSDP v2

## Goal

Support communication where data-parallel ranks contribute and receive unequal
amounts of data, without requiring padding to a uniform size.

## Motivation

TensorAtomic placement can assign different amounts of data to each rank, so
its materialization and gradient-reduction paths need uneven communication.

This is a general communication need, not specific to TensorAtomic. MXFP8
block-atomic sharding may also benefit from uneven communication to reduce
padding overhead.

Separating the work lets the solution be designed around the broader set of
callers and their performance requirements.

## Related issues

- #6368 — Add TensorAtomic DBuffer placement
- #5615 — Support MXFP8 with MFSDP v2 block-atomic sharding