Add serving-llms-on-instinct skill

[Mahdi-CV]

Adds the serving-llms-on-instinct skill: end-to-end LLM inference serving on AMD Instinct GPUs (MI300X/MI325X/MI350X/MI355X) with vLLM on ROCm. The skill handles GPU detection, environment validation, vLLM configuration, launch, and health verification, and refuses non-servable models (diffusion, audio, embeddings, rerankers) with an explanation.

What's included

• SKILL.md and reference.md: skill definition and runtime guidance
• scripts/detect.py: GPU detection via amd-smi (local or remote host)
• scripts/validate.py: environment validation with auto-fix
• scripts/sync_recipes.py: refresh recipes from vLLM recipes + Docker Hub
• scripts/estimate_vram.py: weight + KV-cache VRAM estimation (handles quantized models)
• data/recipes_cache.json: model configs synced from vllm-project/recipes
• data/gpu_overrides.json: GPU-specific docker flags and legacy model configs
• data/blacklist.json: models that cannot be served as LLM endpoints

Registration

• Added the plugin entry to .claude-plugin/marketplace.json and .cursor-plugin/marketplace.json
• Updated the README skills table: serving-llms-on-instinct moved from planned to in-repo

[danielholanda]

PR looks great. Next step here is to add a quick walkthrough so other folks have a bit more guidance when trying this:
#58