Mixed-precision deployment

Ship FP16, BF16, or FP8 training and inference that holds accuracy while capturing the speedup, using loss scaling and numeric validation. Use when moving a model off FP32 to run faster or fit in less memory and the result must stay correct.