Sample and FullyConv Generators perform all sampling/transform logic in flow

[willgraf]

Major Changes

• Simplified data_utils.py, image_generators.py, and training.py.

• Moved all sampling logic from data_utils to image_generators. Now data is sampled on .flow(), and takes in balance_classes and max_class_samples to turn off the class balancing.

• .flow() also takes in a transform key that will transform masks into one of watershed, deepcell, disc, or None (for fg/bg). This transform key replaces all the various other data generators, which have been removed.

• Similarly, all training functions have been removed except train_model_conv and train_model_sample. These two now take in the transform key as well as transform_args which are various arguments for each transform. The use of 2D or 3D generators is inferred from the data in the train functions.

• removed skip flag from the train_model_conv, infer the skip number from model.output_shape

• Notebook are being split into Fully Convolutional and Sample Based notebooks.

• All notebooks have been re-run to reflect these changes.

• Changed deepcell_transform to separate background_eges from cell_edges

[willgraf]

Fixes #72