serve: load GGUF models end-to-end
model_manager only received a model id, so serving a GGUF failed. Auto-resolve a single .gguf (via _maybe_auto_detect_gguf) and pass gguf_file to AutoConfig + from_pretrained and to the processor load, converting a local .gguf path to dir+filename. Fall back to AutoModelForCausalLM when a GGUF-derived config has architectures=None (routes by model_type).
