[Feature Request] Add info dict key-spec pairs to `env.observation_spec` if they're provided to a gym environment through `set_info_dict_reader`

[vmoens]

Motivation

gym's environments can pass some more info through a dictionary info:
obs, reward, done, terminated = env.step(action).
This dictionary can contain additional information about the state of the env: time spent doing the task, whether the task is solved etc.
We wrap the obs, reward and done in the output TensorDict. We can also place the info if they're numerical values (see #234):

        >>> from torchrl.envs.libs.gym import GymWrapper
        >>> from torchrl.envs import default_info_dict_reader
        >>> reader = default_info_dict_reader(["my_info_key"])
        >>> # assuming "some_env-v0" returns a dict with a key "my_info_key"
        >>> env = GymWrapper(gym.make("some_env-v0"))
        >>> env.set_info_dict_reader(info_dict_reader=reader)
        >>> tensordict = env.reset()
        >>> tensordict = env.rand_step(tensordict)
        >>> assert "my_info_key" in tensordict.keys()

Problem is that we also register what are the expected observations (domain, dtype, device etc) in an observation_spec attribute. It is important for us to know what to expect as output from the env. However the info is not yet registered in observation_spec. This is a problem for parallel environments as we would wish to pre-allocate the tensors that we place in share memory to pass information from one process to another: for that, we use the observation_spec which is cheaper than resetting the env.

Solution

We should be able to provide the specs corresponding to the info keys in the default_info_dict_reader, and the set_info_dict_reader should be able to read them. If no spec is provided, a default unidimensional, floating-point unbounded spec should be assumed.