Add parallel features

[dustinvtran]

• asynchronous updates, i.e., HOGWILD
• multiple sgd runs ("chains"): this is easy to do using the doParallel package as it doesn't require any parallel tinkering in C++

[dselivanov]

+1 for HOGWILD - performance very impressive (at least for me). Recently I implemented parallel async adagrad for learning GloVe word embeddings - speed-up is linear. But, to be fair, my problem is quite sparse - feature matrix has ~ 1% non-zero elements.