When switching to the Adam optimizer on Apple computers (as SGD crashes on Apple Silicon), the learning rate was not updated. This often (almost always) caused the models to diverge.

This PR updates the learning rate, to the same values (and schedule) as the default for multi-animal projects (which use Adam):

cfg["multi_step"] = [[1e-4, 7500], [5e-5, 12000], [1e-5, 200000]]

This was tested on an M2 MacBook Air, on a dataset which diverge in less than 5 iterations using the previous schedule, and now converges nicely.