swap: Qwen2.5-0.5B-Instruct (faster, reliable #### format) + fix dp2 detection bug

feat: pre-load model at startup to avoid 200s first-request wait

fix: handle None gold_prob/prod_prob in steps display

fix: context-aware Step token detection for TinyLlama (handle multiple token IDs)

Switch to TinyLlama-1.1B for CPU, remove bitsandbytes

HF Space: Reasoning Trajectory Demo