[Feature Request] Calling `tensordictmodule.forward(...)` with keyword arguments and not tensordict

Motivation

To make TensorDictModule easier to access, a cool feature would be to be able to call the forward method with the tensors and not the tensordict, e.g.

>>> # currently
>>> tdm = TensorDictModule(module, ["a"], ["b"])
>>> tensordict = TensorDict({"a": torch.zeros(3, 4)}, [3])
>>> tdm(tensordict) # populates the "b" key
>>> tensordict2 = tdm(a=torch.zeros(3, 4))
>>> assert (tensordict == tensordict2).all()

Solution

This solution should take the form of a decorator:

@dispatch_kwargs
def forward(self, tensordict):
   do_stuff()

where @dispatch_kwargs will call make_tensordict from the kwargs provided and call the function with the tensordict that has been created.

Additional context

The solution should cover TensorDictSequential and ProbabilisticTensorDict too.
It could also be nice to support sorted inputs (args and not only kwargs).