Proposal outside of KDE returns `-np.inf` which kills sampling

[astrolamb]

If proposal results in a value that is greater than the PDF grid, then likelihood returns -np.inf to effectively kill that proposal.

For strong signals, this can cause unintended, poor sampling, especially if the free spectrum prior is small.

TODO

• Quick fix: warn user when proposal exceeds PDF grid
• Quick fix: replace -np.inf with a very negative but finite number to discourage but not constrain sampling
• More work: when making ceffyl more general, how do we deal with proposals exceeding grid-prior bounds that actually have some support, but KDE was not trained beyond these points

ceffyl/ceffyl/Ceffyl.py

Line 508 in 32f5b43

# else compute probability given data