[Feature Request] Support for nested tensors

[vmoens]

Motivation

Support for nested tensors should work as torch.stack(tensordicts, 0)
(we can't currently override torch.nested.as_nested_tensor so we'll need to write a custom op)

We don't need to subclass TensorDict for this but there are a few caveats:

1. Reshaping won't be allowed
2. Shape can't be accessed
3. Indexing can only be done on the first dim

For instance, populating a tensordict with a nested tensor won't work as we can't access the shape.

• If nested tensors are found in a tensordict, all tensors should have that feature.

• Indexing such a tensordict along the second dimension will require:

1. splitting the tensors that are nested
2. indexing those tensors
3. re-nesting them