add operator overloading to RandomVariable

[dustinvtran]

Many operators are overloaded for RandomVariable, following those available in tf.Tensor. The output is always a tf.Tensor. It operates on the sample tensor associated to the RandomVariable.

For example, this lets us do

from edward.models import Normal

x = Normal(mu=0.0, sigma=1.0)
y = Normal(mu=0.0, sigma=1.0)

x + y, x * y, x / y, x // y
x == y
hash(x)
-x
x ** y

Remarks

• An open question is what to do with the __getitem__ method. That is, should we also return indexing from the sample tensor? Or should __getitem__ be the only method that returns a subset of the random variable? (If so, how do we do so efficiently?)
• Should eval() be a method? E.g., calling x.eval() would be easier than x.value().eval() which is almost always what the user intended. I think it should only be a method if we can also do sess.run(x) somehow, and not require sess.run(x.value()).

todo

• unit test