[partitioning] Support no-collectives local-write scatter along sharded inserted_window_dims.
Support a 0-collective scatter mode when sharded `inserted_window_dims` are
indexed via `scatter_dims_to_operand_dims` (with replicated/compatible indices
and updates, and non-`CLIP` mode): each shard shifts `scatter_indices` by
`-axis_index * shard_size` in `in_collective` and writes only updates whose
shifted indices fall within its local `[0, shard_size)` range, relying on
`hlo.ScatterOp` dropping out-of-bounds updates on all other shards.
This unblocks sharded-axis point/slice updates (e.g. `arr.at[0, ...].add(...,
out_sharding=...)`) as well as `jax.grad` of `top_k`, `approx_top_k`, and
`gather` along sharded axes with 0 communication collectives.
PiperOrigin-RevId: 1024202392