Investigate batch_size usage

[rg936672]

Investigate the strange behaviour of batch_size with values close to the number of training points. With batch_size=num_training_points, training is faster than "unbatched" training, despite the fact that the full dataset is used. With batch_size just barely less than num_training_points, the model seems to make wildly inaccurate predictions.

Strangely, setting batch_size=500 makes the test take only 1.85s, and setting batch_size=499 makes the test fail as 55% of outputs are outside the confidence interval!
- @rg936672 in #354 (comment)