Upgrade Gymnasium version

[TheEimer]

Most of the update was simple, but there are some interesting changes:

1. Final obs are now always shown and then we throw away the next transition (only for updates, not in logging). In practice, this means replacing these transitions with nans, setting them to 0 in the trajectory buffer and not adding them to replay.
2. We can't access the "envs" list in the vector envs anymore
3. Dependencies are a problem: pufferlib and CARL can't be installed with the new numpy/gymnasium version

So this needs a review and we need to decide what to do about CARL and pufferlib. @AndreBiedenkapp might have time to look at a gymnasium update in CARL, but we're powerless for pufferlib, obviously. Pufferlib still works with the new version, though, it's simply an install conflict we could in theory work around. I didn't fix this in the dependencies yet since we should discuss this together.

[TheEimer]

Would close issues #103 and #108