Raise messages if type mismatch when building inference or criticism

[dustinvtran]

Motivated by https://discourse.edwardlib.org/t/poisson-gamma-regression/47.

We can raise an error message, e.g., if a random variable has type float32 but the data passed in has type int32.

We should also do this if there's a difference in floating point precision (for tf.Tensors). For NumPy arrays, we don't need to raise any message and can continue casting under the hood; this is the convention TensorFlow takes.