End-to-end knowledge distillation CLI.

This crate provides a complete pipeline for knowledge distillation:

• Fetch teacher models from HuggingFace
• Configure distillation parameters via YAML
• Train student models with progressive/attention distillation
• Export to SafeTensors, GGUF, or APR formats

Toyota Way Principles

• Jidoka: Pre-flight validation catches errors before expensive training
• Heijunka: Memory estimation enables level scheduling of GPU resources
• Kaizen: Configurable hyperparameters enable continuous improvement