record seed / random state

[alisandra]

record prior to training, so that models are reproducible, and for those rare enough cases where it applies and neither parameter count nor data changes, training can be more consistent when trying to compare more subtle differences.

Anyway, not sure if it will often be helpful in practice, but recording won't hurt.