[spirv] Unify matmul/convolution reduction tiling (#7315)

Previously we have been doing reduction dimension tiling for
matmuls at the time we do tiling and distribution to threads.
This can be confusing because we provide all (M, N, K) tile
sizes there but under the hood we treat them differently:
M/N gotten tiled and distributed, while K is just tiled.

While for convolution, we always only provide configuration
for parallel dimensions for workgroup/thread tiling and
distribution. Tiling for the reduction dimensions is handled
later. It's also inconsistent here.

Therefore this commit moves all reduction dimension tiling
into the same place for clarity and consistency.