Dev refactor for release

[jostosh]

Will contain a significant refactoring for LibSPN release.

Includes the following changes:

• Take out custom Ops
• Fix failing tests
• Take out GradientType since it lacked unit tests and was not giving any empirical improvements
• Use a single DenseSPNGenerator
• Improved lru_cache usage
• Improved docs
• Refactoring for GD Learning
• Take out ValueType and support usage with tf.initializers
• Add type to get_num_nodes
• Add get_depth method
• Move convert_to_layer_nodes to a separate module
• Remove redundant math.py functions (e.g. broadcast_value)
• Add two notebooks for discriminative and generative learning on full-resolution MNIST