Kernel Float: Unlocking Mixed-Precision GPU Programming