Fix lm-eval continuation scoring after partial context truncation

[sqliangx]

Summary

This PR fixes a scoring bug in the lm-eval adapter's MLXLM.loglikelihood() method. Currently, MLXLM.loglikelihood() uses the truncated prefix length to slice untruncated full sequences. As a result, it incorrectly includes context tokens in continuation scoring after partial context truncation.

For example, with a character tokenizer, context abc, continuation de, and max_tokens=3, the context becomes bc. The code should score de, but it currently scores cde instead.

The PR fixes the bug by saving the original continuation boundary before truncating the prefix. This keeps the scoring focused on the continuation. The existing truncation policy and fully truncated (-inf, False) behavior stay the same.

Testing

This PR adds a test covering no truncation and two partial-truncation cases. Both partial-truncation cases fail before the fix and pass afterward.

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: bug finding

[michalk8]

Thanks for spotting this!