[Feature Request] Make transforms stateless

Motivation

TensorDict is a great tool to carry information from one step to another.
Yet, we still use buffers in transforms like CatFrames or ObservationNorm.
We could spare these by using tensors stored in a tensordict.

The only drawback is memory consumption: we may not want to carry over these buffer in the replay buffer etc.
One way to go about that would be to have a prefix in these keys (or even better: store them in a sub-tensordict) that we could discard in env.rollout or collector.rollout.
Something like

>>> env.step()
TensorDict({
    "_transform_buffers": TensorDict({
        "0": TensorDict({
            "loc": torch.Tensor(...)
        }, batch_size=[])
     }, batch_size=[])
}, batch_size=[])

Open questions:

We need to figure out how to transfer some normalization constants from one env to another.
If we do

train_env = make_env()
t = train_env.transform[0] # an ObservationNorm
t.init_stats()  # Presumably: will create a TensorDictPrimer in the env with the normalization constants.
test_env()
? # here we need to figure out how to gather the stats from the other env