Update Colab notebook: 1.5B model, scaled rewards, tuned hyperparameters

Align with Advanced Llama 3.2 GRPO LoRA reference notebook pattern

Fix format_comparison_metrics_html to accept run_comparison() dict directly

Align train.py and Colab notebook with official Unsloth+OpenEnv GRPO patterns

Update metrics format with drift/oversight tracking, add colab training notebook