Physics-Informed Generator-Encoder Adversarial Networks with Latent Space Matching for Stochastic Differential Equations