Multiple strategies to adjust the learning rate

[yiheng]

In this PR, we add two other learning rate adjust strategies.

1. Poly
   adjust the learning rate as:
   init_lr * (1 - iter / maxiter)^beta
   where init_lr is the intial learning rate, iter is the current iteration, maxiter is the maximum iteration and beta is a hyper parameter.
2. Step
   multiply the learning rate with a give factor gamma every N iteration, where the gamma and N are hyper parameters.

These strategies help the model converge to a good accuracy in large batch size.

The implementation is in optim/SGD. And the old lr adust strategies are also move to that place.