Extend test case to reveal bug.

[rpgoldman]

In sample_posterior_predictive, for ObservedRVs we are supposed to generate samples with a shape that matches the shape of the observed. This fails for DensityDist, as illustrated by this expanded test.

[rpgoldman]

See questions at the bottom
Just pushed a fix along the lines @lucianopaz suggested (thanks!): this checks the dimensions of the sample returned in order to check and make sure that it's correct, and raises a TypeError otherwise. This will trap problems with DensityDist not handling observations properly.
It's possible we could fix this by changing the interface to DensityDist to allow users to provide distributions that take a shape argument, but that's for later.

1. Is TypeError the right exception class to use?
2. I have a yucky structure to avoid repeating the shape-checking everywhere. If anyone has a more elegant way to do this, please LMK.

[lucianopaz]

I think that the test case is already addressed by #3554 so the key feature introduced with this PR is the ShapeError.

The idea of having a special ShapeError exception subclass is nice, but I don't think that we can use it in the _draw_value context. Maybe we could raise it inside generate_samples if the broadcast_shape and dist_shape don't broadcast well with each other. I also don't fully understand why you need the Throw exception.

[twiecki]

Closing due to inactivity, feel free to reopen.