Add the ability to extend corpus after training

[ChrisCummins]

e.g. Train on set of kernels X. Train. Add additional set of kernels X'. Train on X'.

[ChrisCummins]

Two possible approaches for this:

1. Just create a new corpus of X + X', and train on that. This is wasteful, because we'll be training a new model every time.
2. Implement a linked list of corpuses. So model points to a corpus, and the corpus contains a reference to the prior corpus. The child corpus derives it's vocabulary from the parent. And the model (already trained on the parent) can be trained on just the child.