Revert "[LinalgExt] Generalize ArgCompareOp in GPUGeneralizeNamedOpsPass" (#23074)

Reverts iree-org/iree#23015

**Reason**

During my process of plumbing arg_compare op along the VectorDistribute
pipeline, I found that working directly with ArgCompareOp (without
converting it to linalg.generic) will be a better solution.

The reason is that VectorDistribute performs two-level tiling:
 
 * Outer level: Partial reduction (PartialReductionOuterReduction)
 * Inner level: Inter-thread reduction (subgroup-wise)

Add PartialReductionOuterReduction support for ArgCompareOp directly, in
addition to the existing PartialReductionOuterParallel (split-reduction)
support. Extending the existing linalg.generic reduction support to
handle arg_reduce semantics would require drastic changes to the
infrastructure (upstream tilinginterface).

Will add the core implementation back when after tiling we find it is
useful for the bufferization + vectorization + vector distribution
phases.