perf(ggml/hip): add batched ROCmFPX MMQ

[weicj]

Summary

Add gfx1151 batch MMQ support for ROCmFP2, ROCmFP3, and ROCmFP4_FAST weights. This keeps DeepSeek V4 Flash batch prefill on the GPU for both dense projections and routed MoE matrices instead of falling through to per-expert dequantization and hipBLAS calls.

Stack: #502 -> #503 -> this PR.

Changes

• Dispatch ROCmFP2, ROCmFP3, and ROCmFP4_FAST through MMQ on gfx1151.
• Add MMQ tile loaders for the packed ROCmFPX formats and preserve their per-half-block scales during the integer WMMA reduction.
• Generate the three ROCmFPX MMQ template instances.

Validation

• The patch applies cleanly to the fix/feat(ds4): ROCmFPX byte-safe decode + fused-decode path + ds4 optimizations #503 base (6c7ee19).
• Built the MMQ sources into dflash_server with ROCm 7.2.4 for gfx1151. The ROCmFP2, ROCmFP3, and ROCmFP4_FAST MMQ instances compile and link successfully.
• Parsed DeepSeek-V4-Flash-ROCMFP2-STRIX.gguf: its 660 ROCmFP4_FAST dense matrices, 86 ROCmFP2 MoE gate/up matrices, and 43 ROCmFP3 MoE down matrices are covered by this dispatch.
• Runtime exact-sequence, 4k/128, 8k/256, and semantic validation on the fix/feat(ds4): ROCmFPX byte-safe decode + fused-decode path + ds4 optimizations #503 lane remain pending a usable gfx1151 host.