[mlir][shape] Use memref of index in shape lowering

Now that we can have a memref of index type, we no longer need to materialize shapes in i64 and then index_cast.

Differential Revision: https://reviews.llvm.org/D84938