Fast and accurate convolution-type smoothed quantile regression.
Implemented using Barzilai-Borwein gradient descent with a Huber
regression warm start. Construct confidence intervals for regression
coefficients using multiplier bootstrap.