Utilizing variational autoencoders in the Bayesian inverse problem of photoacoustic tomography