FMPE iid inference via SDE bridge failing for uniform priors and gauss-correction methods.

[janfb]

Via the new ODE-SDE bridge from FMPE to score-methods we can perform iid inference using FMPE as well.

This works as expected for Gaussian priors and all score-iid variants. However, for uniform priors and all except the fnpe score-iid variant this is failing: The predictor produces NaNs instantly (the initialization seems valid).

I pushed some fixes to the vf tests in #TODO, fixing this for fnpe, but for the others it would need your expertise @manuelgloeckler 🙏

Can you please have a look?

The issue can be reproduced using the following tests, uncommenting the pytest.skip in line 392.

pytest tests/linearGaussian_vector_field_test.py::test_vector_field_iid_inference

[janfb]

While I cannot reproduce this for NPSE locally, remotely during CD, this now seems to happen for NPSE as well. Only for uniform priors.
I adapted the tests to be skipped for NPSE as well.

[manuelgloeckler]

Mhh, yeah that it runs into nans is a bit weird and might be due to the bounded domain. Not sure why this is happening now, but will have a look.

[manuelgloeckler]

Okay, its something more general:

The drift and diffusion around t=1. just explodes and our current save of "1e6" is to big. This will effectively explode the SDE in the beginning to very high values. It eventually recovers quite fast in the standard sampling but this explosion leads to numerical issues in the IID case.