distributed_writer

Pieces of a no-gather distributed HF checkpoint writer.

Every tensor is sent to exactly one owner rank over batched point-to-point sends and written from there, so no rank ever materializes the whole model and no unit is replicated across ranks.