ENH: Vectorize model evaluation in mnist tutorial

[rossbar]

The model is run against the test set at each epoch to evaluate performance. The evaluation step doesn't involve any sequential updates to the model itself, so it can be vectorized. The main benefits of the change are:

1. Demonstrates NumPy features, primarily vectorization and reductions along axes
2. Significantly improves the performance of the training/evaluation loop

This depends on #66 (includes some of the same commits), so marking as draft until #66 is merged.

[rossbar]

Note for reviewers: everything up to cabbdd1 is in #66. 22cc58c contains the relevant code changes if you want to review diffs in the mean time.

[rossbar]

I'm going to cherry-pick these commits over to #68 and close this in the interest of trying to get all of the performance enhancements into a single PR.