Remove redundant compress() in backwards algorithm

[jeromekelleher]

In #2799 we added an initial version of the backwards algorithm, but ended up including two calls to compress() per site. This is because we need to call compress immediately before storing the ValueTransitions, because we rely on the ValueTransitions being listed in tree-preorder for downstream algorithms (like decode()). We also needed to filter out node=-1 values on the first site in the backward matrix (but that's easy enough to solve).

This is surely unnecessary, and there must be a way in which we can avoid having two calls to compress.