This is the first step to convert vector ops to MMA operations in order to
target GPUs tensor core ops. This currently only support simple cases,
transpose and element-wise operation will be added later.
Differential Revision: https://reviews.llvm.org/D102962
