Add default values to constraints to make grad propogate properly

[fehiepsi]

Fixes #871.

In pyro-ppl/pyro#2447, @fritzo proposed to implement a safe_mask function, but it seems to me that this is an easier to implement solution.

[fritzo]

This is a great idea, especially since constraints do the checking.

Could I suggest a different interface, so as to be compatible with PyTorch, where we might need to know precision and device placement?

class Constraint:
    def default_like(self, prototype):
        raise NotImplementedError

class Positive(Constraint):
    def default_like(self, prototype):
        return np.ones(prototype.shape, dtype=prototype.dtype)

This is similar to torch.zeros_like() and similar functions. If you're ok with this, I'd like to copy this pattern in PyTorch/Pyro. We could alternatively name this .feasible_like().

[fehiepsi]

Thanks, @fritzo! I'll update the api. It looks reasonable to me.

[fritzo]

Hmm I think I prefer .feasible_like() over .default_like(), since the former evokes the language of constraints. @fehiepsi do you have a preference?

[fritzo]

Looks great! Mind if I port these to PyTorch, or would you like to?

[fehiepsi]

Thanks for reviewing and valuable feedbacks, @fritzo! I would appreciate if you port those to PyTorch. :)