Support nontrivially shaped bint domains

[fritzo]

Motivation: this is required for e.g. Independent(Categorical) distributions and for Function tensors that input or output batched categoricals.

I believe this is a big change requiring pervasive updates to all shaping logic.

One design question is whether we allow heterogeneous bound as in e.g. heterogeneous Multionmial distributions. Maybe an initial version could implement homogeneous bound, then later a heterogeneous bound.

[fritzo]

We may want to separate Bint from Reals, then generalize Reals to Array with a .dtype property. Some reasons for separating Bint from Array[int] include:

1. Numpy broadcasting semantics can behave differently for ints versus arrays
2. Tuple getitem naturally inputs Bint but not arrays (indeed Bint can be seen as dual to Tuple)

[eb8680]

I'm going to close this as resolved by #417 and #356; if we need to implement specific patterns e.g. Reduce over nonscalar Bint domains for individual terms we can track them in separate issues.