Kernel embedding of maps for sequential Bayesian inference: The variational mapping particle filter