Add mixed precision training and tokenizer lookup optimization

Add accumulation and resumable token-budget scale training

Add validation-gated scaled training and overfitting controls

Add resumable training metrics and role-specific SFT preparation

Add Ares Lab source and Colab training notebook