[mlir][Vector] Don't fully unroll transfer_writes of n-D scalable vectors (#71924)

It is not possible to unroll a scalable vector at compile time. This
currently prevents transfer_writes from being lowered to
arm_sme.tile_writes (downstream).