Another change to SGD/Momentum/Adagrad

[joshday]

I touched a lot of files here, so I wanted to make an issue of it.

I added two fields for all three types: β0::Float64 (intercept/bias term) and intercept::Bool (whether β0 needs to be updated). intercept = true by default.

This solves two things:

• Adds an intercept without making a copy. BiasVector/Matrix is awesome, but there's no rowvec_view() method, so Adagrad(BiasMatrix(x),y) throws an error. There may be a way to fix this.
• The intercept term will no longer be affected by the penalty (as it shouldn't).