NaNs during NLE training

[janfb]

I noticed that NaNs occuring during NLE training, e.g., during evaluation of the loss, are not detected or reported: the training finishes and just returns NaNs all the way. For NPE, we have checks on the loss, e.g., here:

https://github.com/sbi-dev/sbi/blob/9152e933871eee48a40be6f67b9f802a37c41b69/sbi/inference/trainers/npe/npe_c.py#L357C56-L358C1

and we should have a similar check in NLE as well. Or start a discussion whether to move such checks to the ConditionalDensityEstimator level.