Add a more complex learn rate scheduler

[vellusorvari]

Chapter 4 summary, page 126:

Decay your learning rate over time using either linear or exponential decay as shown in this chapter or a more modern technique such as cosine decay. In fact, more effective learning rate schedules vary the learning rate based not just on each epoch but also on the loss on the testing set, decreasing the learning rate only when this loss fails to decrease. You should try implementing this as an exercise!