Raise a NullMask() error for degenerate token masks

[gabegrand]

As discussed in #21, there are various scenarios in which token masking results in a LMTokenMask distribution that has a null support. The previous PR partially solved this problem by returning -inf, which avoided a cryptic ValueError message. However, this still results in an uncaught downstream exception the next time sample() or observe() is called, making it hard to localize the source of the error when debugging.

On further reflection, a complete solution to this problem would be to implement tracing -- minimally, for mask objects, but ideally at the level of the hfppl itself -- to make it possible to track sequences of stochastic events leading to errors. I think this is a potentially useful approach but would require deeper refactors to the library than I have bandwidth for at the moment.

In the spirit of keeping things simple for now, I think it makes sense to introduce a NullMask() error that gets raised when observing a degenerate mask. This has the sole purpose of providing a more descriptive error and a clear stack trace for what I've found to be a common corner case in hfppl programs.