[BUG] NaN when building self.ends

[FactoDeepLearning]

In sparse_hmm.py, in the unpack_edges function, the initialization of self.ends is dangerous as they are no guarantee that all states will be linked to ending Silence(). So, some values may not be initialized, leading to NaN in my case (prefix-tree implementation with SparseHMM).

https://github.com/jmschrei/pomegranate/blob/master/pomegranate/hmm/sparse_hmm.py#L63

Proposed fix:
Replacing:
self.ends = torch.empty(n, dtype=self.dtype, device=self.device) - inf
by:
self.ends = torch.full((n, ), fill_value=-inf, dtype=self.dtype, device=self.device)