Add RDNA4 MoE WMMA kernel path

[vivienfanghuagood]

Summary

• add an RDNA4-specific MoE 2-stage WMMA path for gfx120x/gfx1201
• factor the public MoE compile wrappers so CDNA, gfx1250, and RDNA4 share the same external API shape
• add RDNA4 test coverage, benchmark integration, and user-facing documentation
• make the RDNA4 stage2 fp16/bf16 atomic path stream-safe and cover a real masked-reduce case

Testing

• python3 -m py_compile kernels/rdna_moe_gemm_2stage.py kernels/moe_gemm_2stage_common_gfx1250.py tests/kernels/test_moe_gemm_rdna4.py tests/kernels/test_moe_gemm_wmma_gfx1250.py
• pytest -q tests/kernels/test_moe_gemm_rdna4.py
• pytest -q tests/kernels/test_moe_gemm_rdna4.py::test_moe_reduce_valid_mask_masks_invalid_routes
• sh -n scripts/run_benchmark.sh

Validated on gfx1201 (Radeon AI PRO R9700).