V0.3.3 - Simpler error handling

• Slight change to the TrainingDataGenerator interface but I don't think this is enough to warrant a new minor version
• Added MustXxx functions to the model, which gives the developer a choice to either handle errors correctly or panic upon an error. There are situations where panic is very appropriate (for example, if you know a model should fit some data, any error that has been raised is likely severe, and probably not recoverable, such as having the wrong-shaped input data)
• Added a new example for FitGenerator which fits a model to a sin curve then plots the results