llama-server loads multipart GGUF via models.ini

The introduction of models.ini (via the --models-preset flag) marks a significant step in llama-server's evolution from a single-model endpoint to a robust multi-model router. This is particularly crucial for the latest generation of massive open-weights models like Qwen3.5-122B.

• –Automatically handles split GGUF files (e.g., -00001-of-XXXXX.gguf), removing the need for manual file merging or complex shell scripts.
• –Centralizes model-specific parameters like temperature, top-p, and GPU layer offloading, which previously had to be passed as individual CLI flags.
• –Enables on-demand model loading and eviction, optimizing limited VRAM for developers running several large models simultaneously.
• –Standardizes the user experience for deploying high-bit quants of large models that frequently exceed the 50GB file size limit of many storage systems.