[partitioning] Support replicated-to-sharded broadcast_in_dim.
Slice replicated non-singleton input dimensions to the target output shard via
`dynamic_slice` and `parallel.axis_index` in `_broadcast_in_dim_collective_rule`
when `out_sharding` shards those dimensions, matching `_naryop_collective_rule`.
PiperOrigin-RevId: 1021071866