[mlir] Do not bubble up extract slice when it is rank-reducing.

The bubble up logic was written by assuming the slice operation is
always a normal slice that outputs a tensor with the same rank.

Differential Revision: https://reviews.llvm.org/D124283