Fix batch tokenization for multiple invoke

[cadentj]

Tweaked the indexing on _batch method in LanguageModel.

In this line

batched_inputs["attention_mask"][:1, : attention_mask.shape[1]] = attention_mask

The first [:1] batch index of the re-tokenized batch is set to the running batch's attention mask. This breaks for more than two invokes because attention_mask is shape [n - 1, seq] which is longer than the single index replacement.

Fixed to use negative indexing

batched_inputs["attention_mask"][:-1, : attention_mask.shape[1]] = attention_mask

[JadenFiotto-Kaufman]

@cadentj This doesn't seem to pass the test_invoke test. in test_lm