Variational Bayesian approximation of inverse problems using sparse precision matrices