Megatron

Changes
- Generalized SklearnMetric wrapper to simple Metric wrapper, which can take any function.
- Remove any attempt to create built-in metrics; they exist elsewhere and should be imported from elsewhere.
- Pipeline transform always returns a list of results, even if there's only one output node.
- Renamed parameters of storage.read.

Changes
- Nodes no longer have names, except InputNodes.
- Since nodes don't have names, retrieving columns from cache comes by integer index.
- Input moved from nodes to layers.
- Layertools module added.
- FeatureSet has been removed. Mapping and applying a layer to a set of nodes has been added in layertools module.
- Data readers give dictionaries rather than FeatureSets.
- keep_data flag added to transform to indicate whether to wipe non-output nodes' data or not.
- transform_generator will always keep data.
- Metrics are now their own kind of node, not part of the transformation pipeline; they're run when evaluate() is called.

Changes
- Multiple Keras models no longer possible, but fit_generator is used when fitting to a generator which makes the pipeline load in parallel with GPU training.
- Data index is a string indicating the key name from the passed in data dictionary when using transform.

Changes
- Add wrapper for metrics.
- Add support for multiple epochs for models.
- Fix fitting to a generator so it properly supports models.
- Add pop method to FeatureSet.
- Remove Pandas formatting for output data.
- Import data loaders and generators straight from io module.

Changes
- Allow for mix of Nodes and FeatureSets when calling Layer.
- Re-arrange the layers module slightly (remove common, move to shaping).
- Add internal utility for flattening irregular list of lists.

Bug Fixes
- Name of single unnamed node was coming up as list; fixed.