[mlir][GPU] Move amdgpu target-specific transforms out of generic GPU dialect transforms

AMDGPU-specific lowerings of gpu data movement ops do not belong under the GPU dialect.

Relocate the amdgpu-specific patterns within these files to somewhere more appropriate under `AMDGPU/transforms`:
- https://github.com/llvm/llvm-project/blob/c8fd662bfa83b3cfa1c7a37ffaec8fac9bb36ccf/mlir/lib/Dialect/GPU/Transforms/SubgroupReduceLowering.cpp
- https://github.com/llvm/llvm-project/blob/c8fd662bfa83b3cfa1c7a37ffaec8fac9bb36ccf/mlir/lib/Dialect/GPU/Transforms/PromoteShuffleToAMDGPU.cpp

Start with the AMDGPU-specific patterns in mlir/lib/Dialect/GPU/Transforms/SubgroupReduceLowering.cpp and PromoteShuffleToAMDGPU.cpp, then inspect the existing AMDGPU/transforms organization for an appropriate destination. Relocate those patterns without changing their behavior, and confirm the generic GPU dialect transforms no longer contain the AMDGPU-specific lowerings.

Written by the indexing model from the issue text.