tp loading: use dist.scatter for sharded redistribute
Each rank owns ~num_mappings/world_size mappings via partition_mappings_
across_ranks; for its owned mappings it now pre-slices the materialized
full tensor into one shard per rank and calls dist.scatter, so every
other rank receives only its own shard.
- cluster bandwidth drops from (N-1)*sizeof(full) to (N-1)/N*sizeof(full)
- no more view-keeps-full-tensor-alive bug: each scatter recv buffer is
a fresh allocation sized to the local shard, no clone needed on rx
- replicated params (tp_layer is None) still broadcast — they're small
- world_size==1 path stays fully local, no comms
- ragged shards (non-divisible sharded dim) raise rather than silently
miscompute; every model we ship divides cleanly under tp_plan=auto
