Statistical and computational trade-offs in estimation of sparse principal components