Generalize LocationScale to discrete distributions

[devmotion]

This PR generalizes LocationScale to discrete distributions. This change is motivated by the example in TuringLang/MCMCChains.jl#263: the distribution of the statistic of interest is a PoissonBinomial distribution whose support is scaled with a constant factor. It seems natural to reuse the existing implementation of PoissonBinomial (which potentially can be more efficient than a regular DiscreteNonParametric distribution, as explained in #1285) instead of implementing a custom distribution for this use case.

When I extended the tests, I noticed the surprising defaults for (log)(c)cdf for discrete distributions which lead to bugs of log(c)cdf for DiscreteNonParametric (and potentially other discrete distributions). I fixed the issue for DiscreteNonParametric in this PR but maybe the bugfix should be extracted to a separate PR. In general, I think it would be reasonable to remove the defaults and make (log)(c)cdf consistent with (log)pdf: each distribution has to decide separately, for which inputs the functions are implemented and in particular if real numbers are rounded to integers.

I moved locationscale.jl to src/univariate/ in the first commit, so for reviewing it might be helpful to inspect the first and the other commits separately.

[devmotion]

This generalization would be useful for the example mentioned above (otherwise I would have to copy and slightly modify the implementation of PoissonBinomial) but what is the general opinion here? I guess at least the problem with (log)(c)cdf should be fixed.

[devmotion]

Bump 😃