Reland "[DispatchCreation] Run preprocessing before..."

[IanWood1]

Land #18934 after fixing mi250 regressions.

[IanWood1]

Update: the only difference in the IR after dispatch region creation is the indexing maps. This seems to be an issue where codegen handles transposed input better than transposed output (although they are trivially interchangeable).

• Before pr https://gist.github.com/IanWood1/7ce5d4b713d30a5fc461476e46fbce77
• After pr https://gist.github.com/IanWood1/723897603729c6612ce18e4bff762fff

Also, in these examples, the linalg.generics should be fused with their consumers (conv_2d). This requires bubbling expand_shape through collapse_shape ops. This would probably also resolve the issue