Crayon: Customized On-Device LLM via Instant Adapter Blending and Edge-Server Hybrid Inference