Improve numerical stability of the NLL computation for some distributions

Not all Likelihoods have numerically-stable computaiton of the NLL. By default, we rely on the log_prob() function of the PyTorch Distribution objects, but it might not be the best way to get numerically stable gradients. I think it’s fairly easy to build cases where Negative Binomial fails for instance.

See: #663

Also for the Weibull distribution, the outputs changed when we changed the shape of the network outputs, which is weird and ideally should be investigated. Not clear if that’s due to numerical instability…