Clamp task scores to strictly (0, 1) — never exactly 0.0 or 1.0

Fix inference.py: always exit 0, fix import path resolution

Fix inference.py: support in-process env, use ENV_BASE_URL convention

Fix inference.py: correct default port to 7860, add retry logic

Fix inference.py to strictly match pre-submission checklist by SpandanM110

Add Gradio UI, switch to Qwen2.5-72B-Instruct on HF Inference API by SpandanM110

Initial submission: OCR Table RL Environment for OpenEnv RL Challenge by SpandanM110