[mlir][linalg] Fix generic reduction vectorization

We shouldn't broadcast the original value when doing reduction. Instead
we compute the reduction and then combine it with the original value.

Differential Revision: https://reviews.llvm.org/D111666