[GlobalOpt] Fuse transpose into matmul-looking linalg.generic (#22901)

This PR adds a new pattern `FuseTransposeThroughGenericReduction` that
fuses transpose operations into `linalg.generic` ops with reduction
dimensions by absorbing the transpose into the indexing map. This
pattern will fuse producer transposes with multiple uses (vs other
patterns in this pass that look for single use transposes). Importantly,
we check to make sure that the reduction op's new indexing map is
contiguous, which offsets the transpose having multiple uses.


These changes are needed because `PropagateLinalgTranspose` handles
named matmul ops differently than `linalg.generic` matmuls. It will fuse
transposes with multiple uses into named matmuls, but not
`linalg.generic` ops. CLIP's weights are transposed from NxK to KxN, so
fusing the transposes means that the K dim becomes contiguous. This will
allow us to remove the `--iree-opt-generalize-matmul=false` flag from
CLIP benchmarks without regressions. This was generalized to all
reduction ops because some of the matmuls have an M dim of 1 which means
that after generalization and dropping unit dims, they don't get
identified as matmuls anymore.


These changes were moved from
https://github.com/iree-org/iree/pull/22790

---------

Signed-off-by: raayandhar <rdhar@amd.com>
Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
Co-authored-by: raayandhar <rdhar@amd.com>