MLIR emitters: Vectorize column reductions.

[copybara-service]

MLIR emitters: Vectorize column reductions.

Special thanks to github user lingzhi98 who experimented with this in
openxla/xla#11018.

I tried to make the logic as similar for vectorized and non-vectorized
reductions as I could. The vectorized logic looks like this:

• produce N reduced elements per thread, store the intermediate results in
  a vector V
• loop over the N elements of V, writing each one to shmem
• loop over N elements, reading them from shmem and writing the result to
  global memory