Beyond L1: Faster and Better Sparse Models with skglm