Convert subtensor and subtensor_insert operations to use their
rank-reduced versions to drop unit dimensions.

Differential Revision: https://reviews.llvm.org/D101495