[BugFix] Fix DataLoader batching for ragged lazy stacks

[gtnv]

TLDR

DataLoader batches map datasets through __getitems__. LazyStackedTensorDict inherited TensorDictBase.__getitems__, so batching bypassed its own indexing logic and tried to stack unequal tensors.

Point __getitems__ at the existing lazy-stack __getitem__, keeping ragged batches lazy.

Fixes #1296.

Reproduction

import torch
from torch.utils.data import DataLoader

from tensordict import LazyStackedTensorDict, TensorDict, lazy_stack

dataset = lazy_stack([TensorDict(x=torch.zeros(i)) for i in range(10)])
batches = list(
    DataLoader(dataset, batch_size=4, collate_fn=lambda batch: batch)
)

assert [len(batch) for batch in batches] == [4, 4, 2]
assert all(isinstance(batch, LazyStackedTensorDict) for batch in batches)
assert [len(td["x"]) for batch in batches for td in batch] == list(range(10))

On main, DataLoader raises while stacking the unequal tensors. With this change, all assertions pass.

[vmoens]

Please remove DataLoader references from the entire PR artifact.

It appears in the title, body/reproduction, branch name, commit subject, and [new test](

tensordict/test/tensordict/test_lazy.py

Lines 278 to 288 in 6cc46a5

	def test_dataloader_ragged_lazy_stack(self):
	dataset = lazy_stack([TensorDict(x=torch.zeros(i)) for i in range(10)])
	batches = list(
	torch.utils.data.DataLoader( # noqa: TOR401
	dataset, batch_size=4, collate_fn=lambda batch: batch
	)
	)
	assert [len(batch) for batch in batches] == [4, 4, 2]
	assert all(isinstance(batch, LazyStackedTensorDict) for batch in batches)
	assert [len(td["x"]) for batch in batches for td in batch] == list(range(10))

). Frame this as batched indexing instead, and test __getitems__ directly. Suggested title: [BugFix] Preserve ragged lazy stacks during batched indexing.

The implementation itself is correct so I am approving it but there should not be cross-reference between rl and td in PR/issues unless it provides some context and helps discovery.