Extract models from core

[dcrankshaw]

This pull request implements a few things:

• Removes the need to edit VeloxApplication.scala when implementing a new feature model type (#75). Now, as long as the model implementation is on the classpath, Velox will use reflection to create the model instance. This means that veloxms-core can be used as a dependency to applications that want to use Velox to serve models, rather than having to rebuild the package.
• Moves the existing concrete feature model implementations to a separate module called veloxms-examples.
• Makes the average or "warm-start" user model a broadcast variable that is automatically retrained when doing batch training. It is trained by averaging the labels of the training data across all users and using that as the training dataset (#47).
• Allows you to specify a specific KeystoneML git commit to build against.

[dcrankshaw]

I rebased the changes from #74 into this PR, so that one is no longer needed.

[tomerk]

Great work @dcrankshaw, you've now removed the two biggest blockers to using Velox! Only a few minor points

[tomerk]

LGTM