Environment transform weights not loaded at eval time

I'm willing to contribute a PR to fix this, but I've noticed that eval time, learnable environment transforms don't have their weights loaded. For example, VecNorm is loaded, but the weights are not. VecNorm is also loaded in an unfrozen state. To me, this seems strange as at eval you'd want to have the observations normalized in the same way that was seen in training. Was there a reason for this?

Start by tracing how environment transforms, especially VecNorm, are loaded for evaluation and how their weights and frozen state are handled. Compare this with the training-time transform setup and reproduce evaluation from a trained checkpoint. Done means the learned transform weights are restored and evaluation uses the same observation normalization seen during training.