Single Precision Cholesky

[mathDR]

Hello! I found this paper which implements a single precision Cholesky solve with a double precision refinement (only using symmetric back solves).

Given the monumental difference in compute between single and double precision for a GPU, this seems like something that jax could take advantage of. Is there any appetite for implementing this?

I would love to tackle this, but would need some assistance in targeting the correct area of the codebase to augment.

Any ideas or thoughts would be most welcome!