Sparse standardization

[jolars]

Sparase in-place standardization (centering)

This pull request adds support for in-place standardization for the sparse implementation.

set.seed(1)
fit_sparse <- sgdnet(heart$x, heart$y, family = "binomial")
set.seed(1)
fit_dense <- sgdnet(as.matrix(heart$x), heart$y, family = "binomial")

library(latticeExtra)
c(sparse = plot(fit_sparse), dense = plot(fit_dense))

The implementation of this, however, does not seem as efficient as the one in glmnet and adds quite a bit of overhead whenever the features are already on about the same scale.

Other changes

• The step size (gamma) is now used properly in the update of the intercept.
• Some of the data simulations for the tests now generated more realistically
• Intercepts for the multinomial model are now ensured to sum to zero.