Support for constrained variables

[fehiepsi]

It could be nice to have a simple syntax to define a constrained variable. With this, users can be write code without having to worry about which transforms to use (sigmoid is simple but for many constraints, which transform to use is not trivial), for example

weight = Variable("x", constraint=unit_interval)
# weight = Variable("x").sigmoid()
spn = weight  * dist.Categorical([0.2, 0.8], value="v0").exp() \
    + (1 - weight) * dist.Categorical([0.3, 0.7], value="v1").exp() 

I think we can substitute x to spn as follows:

• For the pattern Variable("x").sigmoid(), we can just substitute an unconstrained value: spn(x=unconstrained_value).
• For the pattern Variable("x", constraint=unit_interval), it makes more sense to substitute a constrained value: spn(x=constrained_value). For optimizers, it would be nice to have a "magical" way to substitute an unconstrained value to spn, for example something like spn(x=unit_interval.transform(unconstrained_value)).

It is not clear to me how useful constrained variables will be in other situations, so I would like to consider this FR as low priority. It is a nice topic to discuss though.