Train a regression model with L1-regularization using Stochastic Gradient Descent.
This solves the l1-regularized least squares regression formulation
f(weights) = 1/2n ||A weights-y||^2^ + regParam ||weights||_1
Here the data matrix has n rows, and the input RDD holds the set of rows of A, each with
its corresponding right hand side label y.
See also the documentation for the precise formulation.