[mlir][vector] Keep scalable dims of the mask in FlattenMultiReduction (#242481)
`FlattenMultiReduction` reshapes the source of a masked
`vector.multi_reduction` with its scalable dims but the mask without
them. For a scalable input the mask `shape_cast` then fails verification
(`vector<1x1x[4]xi1>` to `vector<4x1xi1>`: "different number of scalable
dims"). With the inner-reduction strategy the same mask shows up later
as a `vector.extract` type mismatch. This patch builds the mask type
with the source's scalable dims.
fixes #242183