Precision issues in correctness tests

[frazane]

Some correctness tests fail randomly due to too small tolerance. We need to adapt the max (relative and absolute) tolerance of some tests and possibly document whether some specific calculations suffer from floating point precision issues.

[frazane]

jax-ml/jax#27088

[sallen12]

One possible solution is to define separate tolerances for each backend, and to use this consistently across all tests. The current tolerances in the tests could then be replaced with e.g. "rtol=tol", where tol is the tolerance for the chosen backend.