Alternating direction algorithms for constrained sparse regression: Application to hyperspectral unmixing