[mlir][tensor][sparse] don't drop encoding when infer result type (#91817)

A general question is: is it possible to support hooks here to infer the
encoding? E.g., when the extracted tensor slice is rank-reduced, the
encoding need to be updated accordingly as well.