[BUG] NonTensorStack.reshape gives the wrong batch size for a stack with a zero-size dim #2415

On a NonTensorStack with batch size [3, 0], reshape(1, -1) returns batch size [3, 0]; torch gives [1, 0].

import torch

from tensordict import NonTensorData, NonTensorStack

s = NonTensorStack(*[NonTensorData("a", batch_size=[0]) for _ in range(3)])
print(s.batch_size, s.reshape(1, -1).batch_size, torch.zeros(3, 0).reshape(1, -1).shape)

torch.Size([3, 0]) torch.Size([3, 0]) torch.Size([1, 0])

Expected: torch.Size([1, 0]).

Cause, not traced to the line: NonTensorStack.reshape (tensordict/tensorclass.py:6396) reshapes a grid of member positions, which is empty here, and regroups the stack with _view or _select_positions; one of them keeps the old outer shape when there are no positions.

Found while reviewing #2410, which sends view of a zero-size lazy stack to reshape; for NonTensorStack it makes view give the same wrong sizes.

tensordict main (924b72a), torch 2.14.1, Python 3.12.12.