Introduces tests for torchrl.envs.transforms classes.
Together with this PR, we also change the logic by which envs return observations.

Before

Each env.step returned a tensordict with various keys. All the new observations were labelled with the prefix'next_observation'.
The envs observation_spec attribute could have been any type of TensorSpec.

Now

We expect all observation_spec to be of the Composite type. This ensures that all observations are signed with their keyword.
It is not expected anymore that all observations have the 'next_observation' prefix, but each key of the Composite TensorSpec must start with 'next_'. Therefore, the signature of the env step is consistent: the observation spec points to 'next_' keys that can be found when calling step().
For reset, while it is still the default behaviour that the 'next_' is removed, it can be recovered by setting the execute_step to False. This is used by _BatchedEnv subclasses and by the TransformedEnv class to make sure that all observations are marked with a next_ (which is necessary for the transforms to be applied, for instance).