JSON-formatted input for training configuration

[rouson]

This pull request adds

• A hyperparameters_t type encapsulating mini-batch size, learning rate, and type of optimizer
• A network_configuration_t type encapsulating nodes per layer, activation function, and whether or not to use skip connections
• A training_configuration_t type that extends the [sourcery] file_t type and aggregates hyperparameters_t and network_configuration_t objects.
• to_json and from_json functions that create or consume JSON-formatted lines stored in string_tarrays for each of the above types.

Each from_json function can be invoked via a generic interface to a user-defined structure constructor function named after the corresponding type.