Fix: remove assistant prefill, use simple robust approach

Force full reasoning chain with prefilled assistant start

Fix: stronger prompting to force structured reasoning output

Fix: handle streaming response format correctly

Add HF_TOKEN auth + fallback model for robustness

Switch to Qwen2.5-3B-Instruct (free serverless, no auth needed)

Fix: Use HF_TOKEN from environment for Inference API authentication

Fix Gradio 5.31 ChatInterface API (remove deprecated btn args)

Deploy DeepMed-R1 demo Space