Learning of state-space models with highly informative observations: a tempered Sequential Monte Carlo solution