Finding a normal net that is runtime-matched to PonderNet

[stfwn]

1. Run PonderNet experiment.
2. Observe mean number of ponder steps on training set.
3. Run experiment with a normal network that contains this mean times the number of parameters in the PonderNet.
4. Observe that the runtime per training epoch is approximately the same or adjust the normal net so that this is the case.

(5. Compare test accuracies between these two nets to evaluate whether pondering is 'worth it'.)