separate brushfire-tree from brushfire-training

[avibryant]

Some applications, like model servers, will not need the training code. It would be nice if they could depend on a smaller subset of the code. Likely, this would include:

• Tree
• Predicate
• Voter
• the serialization Injections
• Dispatched

but hopefully would not include:

• Instance
• Splitter
• Error
• Sampler

[tixxit]

Something to think about while doing this, is adding a case class Forest(trees: Map[Int, Tree[K, V, T]) in brushfire-training while we're at it. A lot of helper methods could be added to this to make the creation of new types of Trainers easier.

[avibryant]

Forest is probably useful in brushfire-trees too, right? Then I guess we could have an implicit ForestTrainingHelpers...