[EPIC] Support lower precision data types #197
Description
This is a tracking issue for supporting lower precision data types. We should strive to support as many operators on the types as we can where there's appropriate implementations for us to bind / lower to. We should not try to build our own implementations / emulation when there isn't something for us to reasonably lower to.
We should strive to support the intersection of the types between https://github.com/jax-ml/ml_dtypes and https://docs.nvidia.com/cuda/cuda-math-api/index.html and as many of the operators as we can.