[DTensor Bugfix] Explicitly specify grad_placements in to_local to ensure necessary all reduce takes place

[acisseJZhong]

Solved #2217 by @alpemreacar

Problem:

x.to_local() was called on a multi-dimensional DTensor (e.g., on a (dp, tp) mesh with placements like (Shard(0), Replicate())). The bare to_local() strips all mesh dimensions and loses gradient placement info for the non-DP (TP) dimensions. For a Replicate() RMSNorm weight on the TP mesh, the backward should produce Partial() gradients (requiring all-reduce), but this information was lost.

Fix:

1. Extract the non-DP placements from x
2. Compute the corresponding grad placements — Replicate → Partial(), others stay as-is
3. Pass the full grad placements (DP dims + non-DP dims) to to_local()

This ensures that in the backward pass, gradients flowing back through to_local() are properly wrapped as a DTensor with Partial() on the TP mesh dimension, which will trigger the necessary all-reduce across TP ranks.

Authored with Claude.