Kernel may not be positive semi-definite for labelled subtrees

[ArtPoon]

Pending minimal repro case.

[ArtPoon]

When matching subtrees, we currently ladderize the trees so that branching skews to the same side, and rotate cherries so that branch lengths are the most concordant (but do we still do this?).

When comparing subtrees with labels, it is no longer clear how to handle the case where branches are discordant with respect to BOTH lengths and labels.

[ArtPoon]

A consequence of this is that the diagonal entries --- k(t1,t1) --- are not 1 when we normalize all entries by sqrt(k(t1,t1) * k(t2,t2)), where t2 = t1.
Again, I haven't gotten around to building a minimal repro case for this issue yet.

[ArtPoon]

Could be related to #134

[ArtPoon]

This should be addressed in our present refactoring of the kernel, where labels are evaluated as productions at internal nodes instead of at the tips.