Empty TensorDict population in loops

[vmoens]

Allows for TensorDict creation like this

>>> td = TensorDict({}, batch_size=[100])
>>> 
>>> for i in range(100):
>>>     td[i] = TensorDict({"a": torch.randn(10), "b": torch.zeros(1)}, [])
>>> 
>>> print(td)
TensorDict(
    fields={
        a: Tensor(torch.Size([100, 10]), dtype=torch.float32),
        b: Tensor(torch.Size([100, 1]), dtype=torch.float32)},
    batch_size=torch.Size([100]),
    device=cpu,
    is_shared=False)

The tensordict is populated with zero-valued key-value pairs when first initialized, hence creating an empty tensor that is to be populated.

cc @BoboBananas