Suggestion: Make optimisers stateful

[ablaom]

Sometimes you want to train for n iterations, inspect results and then add additional iterations if required. Or one may want to control iteration externally, using a package like IterationControl.jl, which provides more options for early stopping, callbacks, and so forth. This is not possible at present because optimiser "state" (such as the trust region radius Δ) is lost as soon as solve! returns. So, calling solve! a second time means re-initializing state, which is undesirable.

Perhaps optimisers could implement a pattern similar to that adopted in Optimisers.jl for gradient descent updates?