Add mixed precision training and tokenizer lookup optimization

Add Ares Lab source and Colab training notebook