Learning sparsity-promoting regularizers for linear inverse problems