Enable grouped convolutions for CudnnConvBackwardInput.

So far, for grouped convolutions we always use forward convolution, which means
we can't "fuse" the reverse into the cuDNN call. With this CL, we can also allow
to use grouped convolutions if we match the backward convolution case. To make
this work, we need to insert another reshape op.
Also, refactor the code so that it returns the new "rhs" operand.

PiperOrigin-RevId: 212256924