Make loading features more efficient

[SkepticRaven]

Right now loading of features has a lot of room for performance improvements. Larger projects can take more time reading in features than training the classifiers takes, particularly if long videos are used with very sparse labels.

Brief rundown of current strategy:

1. Load labels from project
2. Initialize feature object, which loads all features in the file
3. Load selected features from feature object here
4. Feature loading accesses features through feature_module.items().
5. Subset features to only labeled frames (unnecessary? should already be subset above)

Proposed change:
Option 1:
Change initialization strategy to delay reading in or calculating features until its needed. There's a lot of stuff that accesses the feature functions and we would need to check testing coverage.
Option 2:
Make a static function for IdentityFeatures that will read in selectively. We would need to be careful about handling the situation where the data may not be cached, which is a bit more difficult for window features.