Gpustack

Fix distributing TPS slowdown caused by the RPC server register position.

1. Compile Linux Arm64 with armv8.2-a;
2. Fix failed to distribute offload.

Support CUDA Linux Arm64

1. Support MiniCPM-O Vision
2. Support Deepseek-R1

1. Fix crash on truncating long kv cache;
2. Fix crash on chatting with the image which has Alpha channel;
3. Fix VRAM occupations when zero offloading with `--mmproj`;
4. Compatible with some GGUF files which described the wrong `kv_count`, e.g: [CompendiumLabs/bge-large-zh-v1.5-gguf/FP16](https://huggingface.co/CompendiumLabs/bge-large-zh-v1.5-gguf/tree/main?show_file_info=bge-large-zh-v1.5-f16.gguf).

1. Remove the VRAM occupation when zero offloading: `-ngl 0`;
2. Fix rerank model loading error: [gpustack/gte-multilingual-reranker-base-GGUF](https://huggingface.co/gpustack/gte-multilingual-reranker-base-GGUF), [gpustack/jina-reranker-v2-base-multilingual-GGUF](https://huggingface.co/gpustack/jina-reranker-v2-base-multilingual-GGUF)
3. Support tool calling in ChatGLM4 series;
4. Introduce DDIM(`ddim_trailing`) sample method;
5. Support multiple devices offloading image model.