llama.cpp v0.4.1

## Overview llama.cpp 0.4.1 adds Maple 20B-A1B, Tencent Hy 4, and Spark2.5 support, improves JSON schema handling, chat parsing, logging, and server child-process management, and updates ggml to v0.24.0. ### API changes - Changed to return instead of ([#28631](https://github.com/ggml-org/llama.cpp/pull/28631)). - Added and for router child-process monitoring in ([#28555](https://github.com/ggml-org/llama.cpp/pull/28555)). ### New models - Added Maple 20B-A1B ternary MoE architecture (CPU) ([#27000](https://github.com/ggml-org/llama.cpp/pull/27000)). - Added Tencent Hy 4 () preview architecture support ([#28127](https://github.com/ggml-org/llama.cpp/pull/28127)). - Added Spark2.5 support ([#27868](https://github.com/ggml-org/llama.cpp/pull/27868)). ### Core changes - Added Kimi-K3 recurrent-state rollback support ([#28466](https://github.com/ggml-org/llama.cpp/pull/28466)). - Fixed MTP context KV cache allocation for DeepSeek2, GLM-MoE, and related models ([#28630](https://github.com/ggml-org/llama.cpp/pull/28630)). - Fixed GDN normalization from to for affected Qwen/Kimi/GLM models ([#28068](https://github.com/ggml-org/llama.cpp/pull/28068)). - Fixed Granite parameter counts and Granite3 MoE unknown parameter counts ([#28643](https://github.com/ggml-org/llama.cpp/pull/28643), [#28632](https://github.com/ggml-org/llama.cpp/pull/28632)). - Fixed MIMO2 SWA pattern loading and incorrect uses ([#28865](https://github.com/ggml-org/llama.cpp/pull/28865), [#28868](https://github.com/ggml-org/llama.cpp/pull/28868)). - Added conversion flag to fuse Q/K/V tensors during HF-to-GGUF conversion ([#22780](https://github.com/ggml-org/llama.cpp/pull/22780)). - Wrote explicit metadata…