Widen options for momentum for our APGD class

[MargaretDuff]

Description

For general convex cases (non-smooth, inexact regularisers), we cannot use the current implementation of FISTA combined with stochastic estimators. For strongly convex objective we can run FISTA+SVRG but we need another choice of momentum.

As suggested by @epapoutsellis, we could change the update step in APGD to

Where momentum is like a step_size_method or callback and has access to the whole algorithm class.

To get FISTA you would set the step size method to 1/L and choose the momentum based on nesterov. A constant momentum would also be useful.