Implementation of algorithm one from the paper

[rcmalli]

This PR is the initial effort for implementing Algorithm one for online learning using Warpgrad. I started analysing the implementation of algorithm 2. Since online learning algorithm does not require to store datapoints and model states in the buffer, I have reused step function from warpgrad.utils inside inner training loop.

Summary of changes:

• New wrapper for online algorithm added. This reuses functions from warpgrad.utils
• Simple updater class is added. However, it works only as placeholder and does nothing in the backward pass call. I am not sure if leap based initialization should be applied also for online learning.
• step function is called inside run_batches function of the wrapper class for each k times of inner update.
• Generated losses are accumulated using meta_loss property of wrapper class.