BiBoMoe

Enhance CodeRabbit config with BiBo-specific knowledge and experimental workflow

docs: clarify router mechanisms - logit norm (confidence) vs bias updates (load balancing)

BiBo+SSMax validation complete: 34.8% faster training, 7.6% better loss, 41.5% better PPL vs Qwen3MoE

better lr

Add wandb/ to .gitignore and remove from tracking