[mlir][linalg] Fold fill -> tensor_reshape chain

For such op chains, we can create new linalg.fill ops
with the result type of the linalg.tensor_reshape op.

Differential Revision: https://reviews.llvm.org/D99116