Support for TruncatedDistribution

[fehiepsi]

Currently, we have one-side truncated normal and cauchy implemented. But it would be nice to add a general class, which uses cdf method, for other distributions. There is a nice discussion here for context on relevant numerical issues.

cc @fritzo

[fritzo]

Note we also have the one-sided general FoldedDistribution in Pyro, which is a cheap rsample-friendly qualitatively similar alternative to TruncatedDistribution. This doesn't get us any closer to literally translating Stan models, but maybe we can take license to approximately translate models and e.g. check that we're only slightly truncating, or not truncating past the median (since Folded and Truncated differ for severely truncated distributions).