Add epsilon-size override for ONNX exporter

[tgolsson]

          I'll make a note to augment the ONNX exporter so it can handle this; this feels like it could end up in other places too.

    def forward(self, obs: Tensor, epsilon: Tensor = None):
        # we need to discard the extra dimensions of epsilon.
        # the input epsilon is given for the original action space
        # however, the policy outputs latent actions.
        if epsilon is not None:
            epsilon = epsilon[:, : self.latent_size]

Originally posted by @tgolsson in #159 (comment)