Quantized KV cache fixes merged into DeepSeek V4 llama.cpp branch

Community developer merges fixes for quantized KV cache into a llama.cpp fork supporting DeepSeek V4, including PRs #25247 and #25303. The branch is available on GitHub.