[Codegen][GPU] Add DistributeArgCompare pattern (#23793)

This PR adds the `DistributeArgCompare `pattern to distribute
`iree_vector_ext.arg_compare` operations across GPU threads and
subgroups.
 
For supported comparators, we use a ballot-based approach that leverages
`gpu.subgroup_reduce` + `gpu.ballot` for reduction.
Supported comparators include:
1. Direct comparison on values (e.g., arith.cmpf ogt for argmax)
2. Same unary op applied to both arguments before comparison (e.g.,
math.absf for argmax of absolute values)

Unsupported comparators fall back to the portable butterfly shuffle
approach. Currently, this is mainly used for argmax/argmin operations,
but we can extend support for additional comparators as needed.

Issue: #23005
Assisted-by:  [Claude Code](https://claude.ai/code)

---------

Signed-off-by: Bangtian Liu <liubangtian@gmail.com>