feat: implement Unsloth GRPO training pipeline with environment-backed reward functions and balanced dataset generation

fix: disable AMP for quantized models to avoid gradient scaler issues in GRPO training

feat: update GRPO training configuration with additional parameters for logging and precision

Adjust logging configuration for training: log every step, enable completion metrics, and limit completions printed per step.

feat: implement GridMind-RL training pipeline with GRPO Colab notebook and Unsloth configuration script

feat: implement Unsloth GRPO training script with environment-based reward tracking and balanced dataset generation

feat: add GridMind GRPO training environment and Unsloth training script

feat: add script to migrate max_new_tokens from GRPOConfig to GRPOTrainer in notebook

fix: update training script with seed variation, fix reward normalization, regenerate training curves showing 0.52->0.67 improvement

fix: training reward uses 8-step rollout + /grade for genuine episode-level signal

feat: add baseline evaluation tools and demo scripts for RL performance comparison