feat(ggml): add ROCmFPX quant support for DS4

[davide221]

Summary

• add ROCmFP4/ROCmFPX ggml quant types and CPU reference conversions
• wire ROCmFPX CUDA/HIP dequant, copy, getrows, MMVQ vecdot, MMVF, unary, and FA paths
• preserve existing luce verify-kernel optimizations: graph stats, LUCE_Q8_MEMO, LUCE_MMVQ_MAX_NCOLS, and richer CUDA op diagnostics
• add DS4 SWIGLU parameter plumbing for fused FFN matvec paths

Used By

• Hub draft PR .NET LLaMA.cpp library! 🦙 ggml-org/llama.cpp#494 points server/deps/llama.cpp at this commit: fix(ds4): keep ROCmFPX generation byte-safe lucebox-hub#494

Verification

• git diff --check
• source audit confirmed ROCmFPX type coverage in ggml type registry, CUDA/HIP dequant/copy/getrows/MMVQ paths, and preserved ac06 perf hooks
• lucebox2 HIP release build succeeded as the .NET LLaMA.cpp library! 🦙 ggml-org/llama.cpp#494 submodule stack: cmake --build build-hip-pr494 -j 8 --target dflash_server
• ROCmFPX DS4 model loaded from /home/lucebox2/models/DeepSeek-V4-Flash-ROCMFP2-STRIX.gguf under the updated hub build
• n_gen=64 warm benchmark with DFLASH_DS4_FFN_RAW_MMID=1 and DFLASH_DS4_HC_DIRECT_NO_SYNC=1: decode tok/s aggregate 21.85, range 16.16 - 23.17

Notes

• Draft while reviewers inspect the imported ROCmFPX format code and decide whether to keep all enum values/ftypes exposed in ggml.

[davide221]

Superseded by Luce-Org/lucebox-hub#502, which ports this ROCmFPX ggml work into the hub's vendored server/deps/llama.cpp tree (ggml is no longer a submodule after lucebox-hub#486). The port is layered on current hub main so main-side ggml work is preserved, adds two review fixes (ftype-switch -Wswitch coverage; FP6 vecdot over-read), and was verified on lucebox2 (clean HIP build, decode 22.8-22.9 tok/s).