Adding support for learning rate schedulers

[gallego-posada]

Context: Many current training pipelines involve tweaking the step size/learning rate throughout training to achieve good performance. Two common scenarios for this are:

• Decreasing the learning rate can help to get closer to the optimum, rather than keep bouncing around it.
• It can be useful when considering problems in which the theoretical guarantees involved an assumption on diminishing step sizes.

Proposal: Add capabilities to use Pytorch's LR schedulers along with Cooper.

Challenges:

• It should be relatively straightforward to handle this for the primal_optimizer as it is "fully instantiated" by the user before creating the ConstrainedOptimizer.
• It gets a bit tricky with the dual_optimizer as we don't have access to the full optimizer before the Lagrange multipliers have been initialized.
• One important thing here is to keep the definition of the lr scheduler as consistent as possible with that of the dual optimizer: since we ask for a "partially instantiated optimizer" maybe we should ask for a "partially instantiated scheduler"?
• Shall we just leave it under the user's control (just as in Pytorch) to perform the calls to lr_scheduler.step()?

[gallego-posada]

Closed by #18