[BUG] RandomCropTensorDict never samples the last window and fails when the length equals sub_seq_len #4577

Describe the bug

RandomCropTensorDict samples the window start with torch.randint(max_idx_0, ...), which draws from [0, max_idx_0). The last valid window is never sampled, and when the sequence length equals sub_seq_len the call fails, although the length check accepts that case. The masked path has the same off-by-one: floor(rand * (L - sub_seq_len)) never reaches L - sub_seq_len.

To Reproduce

import torch
from tensordict import TensorDict
from torchrl.envs.transforms import RandomCropTensorDict

td = TensorDict({"a": torch.arange(4).expand(1000, 4)}, [1000, 4])
print(set(RandomCropTensorDict(2)(td)["a"][:, 0].tolist()))  # {0, 1}, never 2
RandomCropTensorDict(4)(td)  # RuntimeError: random_ expects 'from' to be less than 'to'

Expected behavior

Window starts are drawn uniformly from [0, L - sub_seq_len], both ends included, with and without mask_key.

Versions

torchrl main (264a3c2), tensordict 0.14.2, torch 2.14.1 (CPU), Python 3.12.

I have a fix with tests ready and will open a PR.