Fix potential off-by-one error in attention mask generation

[dibyaghosh]

See #75

TL;DR: There is a small bug in the attention masking code; this should not practically affect anyone using the released model or training their own models (unless you're doing some special attention mask scheme), but we will fix it soon in an update.

The issue: If you have multiple timestep groups, the bug causes the first token in the second group to be misclassified as being in the first group (similarly, 1st token of 3rd group is misclassified as being in group 2, so on). If your model relies on different timestep groups not being able to attend to each other (this is a pretty non-standard case), then this could cause undesired information leakage.

For most people (if you are using the released model checkpoints, if you are using our config for pretraining), it should not affect any of your use cases. There might be some weird behavior if you try specifying readouts to a non-standard value in

octo/octo/model/octo_module.py

Line 91 in bd930f9

readouts: Optional[Sequence[str]] = None,