[mlir] Fix folding for scf.for(tensor.cast).

We should only fold tensor.casts that provide some new static information about
shapes, instead of looking for a symmetric pattern cast(for(cast)).

Differential Revision: https://reviews.llvm.org/D144577