Simplify lambda path construction: create it one-time evenly spaced in log-space.
New way of adding strong variables: initialization always puts only the non-penalized variables. These variables are fitted with lasso with lambda = inf. This gives the correct state initialization for lambda_max, which is the smallest lambda such that all penalized variables have coeffs = 0 with the current values for non-penalized coeffs. Subsequent fitting is the same as before.
Added the default option of using strong rule. Old incremental method is still there and is minimally used as a fall-back even if strong rule is used.