logp of Bound distributions is wrong

[aseyboldt]

This came up in #1833.
The logp function of Bounded returns the raw values of the underlying distribution, but it should correct those with the new normalization constant of the bounded distribution using the cdf and ccdf of the underlying distribution. In many cases this is just a constant, so the results of the samplers do not change, but if Bound is used as an observed var (to model truncated data for example), this can lead to wrong results. This should either me mentioned in the doc of Bound, and in this case it should probably refuse the observed keyword, or we would need to implement cdf and ccdfs for each distribution in pymc3.

[domenzain]

The second alternative seems more natural to the users and provides much more modelling power out of the box.

[ColCarroll]

I agree the second seems like the "right" answer, but it might be better for everyone to update the documentation and refuse the observed keyword for now and leave this issue open for someone to take up. It just seems like updating the docs is a ~30min exercise and the other would require some more careful work and testing.

[aseyboldt]

We might also want to somehow keep the functionality of the current Bound around. This does seem useful when defining new distributions.
I think I have a bit of time to work on this sometime this week. I guess we don't need cdfs for all distributions for now.

[domenzain]

If there is a place to gradually add cdfs, I can contribute a few of them too during the week.

[domenzain]

Maybe this could be called Truncated. This way Bounded is kept around for other uses and the current code can stay as it is.

[aseyboldt]

Closed by #1942, see #1864 for long term plan.