Scalable Variational auto-encoders with complex distributions Application to Earth observation representation learning