[sharktuner] sync with IREE's heuristics to take all dimensions into account for tile distribution.

[bangtianliu]

It seems that the tuner needs to catch up with IREE's logic change.

The relevant PRs (as far as I know):

[Codegen][GPU] Adding new heuristics to take all dimensions into account when distributing tiles
[VectorDistribute] Allow distributing subgroups on multiple m dimensions
[VectorDistribute] Use subgroup_basis instead of subgroup_m/n_count

[bangtianliu]

cc @kuhar @Max191 @RattataKing

[bangtianliu]

Also Need to have a look at TileAndFuse pipeline per this PR:
[Codegen][GPU] Enable TileAndFuse for matmul by default

[kuhar]

From discord:

@bangtianliu
I just had a chat with Kunwar. His thinking is that no urgency to sync the task of distributing to multiple m dimension from the PR iree-org/iree#22000 . Focusing on inner most dimension is enough right now for tuner.
The priority should be distributing to inner most dimension (that is what we have now), then if inner most dimensions are small, we focus on distributing to other dimensions.