A folder of tensor_load + tensor_to_memref exists but it only applies when
source and destination memref types are the same.
This revision adds a canonicalize tensor_load + tensor_to_memref to memref_cast
when type mismatches prevent folding to kick in.