[Mosaic GPU] Fix handling of sliced/transposed types in `transformed_smem_ref_type`.
Previously, we would fail to produce correct strides and offsets in cases
involving slicing, leading to verifier errors. Concretely, slicing does not
affect strides---but we would recompute the strides solely based on the
sliced row-major logical shape of the ref.
Contrarily to lane lowering, where tiling transforms are always
applied prior to slicing or transposition when lowering, warpgroup lowering
works with sliced or transposed refs, and must apply tiling transforms on them
a posteriori.
Our function is now able to tile arbitrary memref types.
PiperOrigin-RevId: 863479995