Sparse Bayesian Lasso via a Variable-Coefficient $\ell_1$ Penalty