[VectorDistribute] Distribute `transfer_write` to single write

[sommerlukas]

#23072 adds an alternative vector distribution pattern for vector.transfer_read which effectively expands the shape of source memref to include the entire distributed shape based on the vector layout. Through this reshape, it is then possible to load the entire distributed vector (batch x outer x element) with a single vector.transfer_read from the reshaped memref.

This shortens the IR, improves readability and defers the lowering of the read to the actual lowering of the transfer_read operation.

Analog to transfer_read, the same principle can be applied to transfer_write, which is the core of this task.