perf: fp16 half() for MiniLM embeddings

Switch to gliner2 package for fastino/gliner2-base-v1

Clean Gradio app: extract_entities, embed_text, health via Blocks API

Minimal test: GPU function only

Move @spaces.GPU decorators to UI functions only

Add @spaces.GPU decorators for ZeroGPU support

Switch GLiNER to fastino/gliner2-base-v1 (PyTorch/GPU), drop onnxruntime

Add autolinks-models Space: FastAPI /extract /embed /health + Gradio UI