[2/N] add cuda imatrix support for custom RL model

[yiakwy-xpu-ml-framework-team]

Introduction

Previously we imatrix was assumed to be tuned on Metal machine (yes we have one M3/2 Ultra) but it is handy to tune the file directly on Hopper platform to avoid network traffices for a 250 GB model.

This is a follow up of #368

Usage:

DS4_CUDA_IMATRIX_GPU_COLLECT=1 ./ds4 --backend cuda
-m gguf/DeepSeek-V4-Flash-Q4KExperts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-imatrix.gguf
--backend cuda
--imatrix-dataset gguf-tools/imatrix/dataset/rendered_prompts.txt
--imatrix-out gguf/DeepSeek-V4-Flash-chat-v2-routed-moe-ds4.dat \

Note DS4_CUDA_IMATRIX_GPU_COLLECT=0 , means we will use legacy imatrix logics without gpu acceleration to compute imatrix scores.

[yiakwy-xpu-ml-framework-team]

@antirez I guess this is a useful feature , wish your feedback!