[mlir][Vector] Add vector bitwidth target to xfer op flattening (#81966)

This PR adds an optional bitwidth parameter to the vector xfer op
flattening transformation so that the flattening doesn't happen if the
trailing dimension of the read/writen vector is larger than this
bitwidth (i.e., we are already able to fill at least one vector register
with that size).