Fix dashboard logging URL to use proxy path, force Docker rebuild

Fix MAX_SEQ_LENGTH: 1024 was too small for prompt+completion, bump to 2048

Fix Qwen3 thinking mode: add /no_think, increase max_completion_length

Fix API response structure: done/reward are top-level, not in observation

Hardcode task definitions in notebook to avoid /tasks API dependency

Redesign frontend as training dashboard + add live activity feed

Route rewards through OpenEnv API instead of local computation

Fix GRPO: remove SFT, multi-task dataset, instruct model only

Fix training: add SFT warmup + switch to instruct model

Update training notebook: vLLM fast inference, Qwen3-4B, max_steps=300

Add GRPO training Colab notebook

Add RL training environment with OpenEnv backend