[mlir] Improve/fix fusion in the presence of scalable tiles

This is a meta-task to capture the effort to:
* Enable post-tiling fusion in the presence of scalable tile sizes, especially for `linalg.pack`, `linalg.unpack` and `linalg.mmt4d`.
* Avoid requiring pack/unpack/mmt4d to be split across multiple functions in this PR: https://github.com/llvm/llvm-project/pull/157815 (while keeping tiling and vectorisation enabled).

See also https://github.com/iree-org/iree/issues/21475.

CC @egebeysel