raise more informative error message if random variable does not have sampling

[dustinvtran]

when trying to instantiate a random variable that does not have sampling, we get

>>> x = Poisson(lam=5.0)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "edward/models/random_variable.py", line 80, in __init__
    self._value = self.sample()
  File "/Users/dvt/Envs/venv/lib/python2.7/site-packages/tensorflow/contrib/distributions/python/ops/distribution.py", line 374, in sample
    samples = self.sample_n(total, seed)
  File "/Users/dvt/Envs/venv/lib/python2.7/site-packages/tensorflow/contrib/distributions/python/ops/distribution.py", line 397, in sample_n
    self._check_hasattr(self._sample_n)
AttributeError: 'Poisson' object has no attribute '_sample_n'

this error is a huge point of confusion for our users. we should catch this error and raise our own with a more informative message.

[laufere]

If I understand correctly (since as a user I was hugely confused :)) that means that edward does not support Poisson as of yet? But it is still possible to instantiate a Poisson random variable since tensorflow provides it as a distribution (with le pmf or logpmf being the minimal requirement to be a valid distribution)?

[dustinvtran]

I recommend looking at the Advanced Settings in http://edwardlib.org/api/model-development. It shows how you can implement your own sampling if desired, or how to instantiate a Poisson just as a likelihood if you don't want to implement sampling.