[mlir][std] Fold load(tensor_to_memref) into extract_element

This canonicalization is useful to resolve loads into scalar values when
doing partial bufferization.

Differential Revision: https://reviews.llvm.org/D91855