Extend forest kernel interface

[andrewherren]

To support a wider range of statistical methods that construct and work with kernel representations of ensembles (i.e. by defining a distance metric based on observations' shared leaf membership), we need to return:

1. The largest possible leaf index in an ensemble (as a complement to the observation-specific leaf indices returned by compute_leaf_indices
2. Observation-specific leaf node variance parameters for forest that support this parameterization (i.e. univariate Gaussian leaf models)

[andrewherren]

Addressed by PRs #99 and #100