User merges quantized KV cache fixes into DeepSeek V4 branch

Fairydreaming merged fixes for quantized KV cache (PRs #25247, #25303) into their DeepSeek V4 branch of llama.cpp. The branch is available at github.com/fairydreaming/llama.cpp/tree/dsv4.