fix: compatibility with mlx_lm 0.31.0 prompt checkpoints

[rsnow]

Closes #110

mlx_lm 0.31.0 added prompt checkpoint support to BatchGenerator, changing insert() tuple arity from 6 to 7 fields and replacing hardcoded prefill boundaries with a variable prompt_checkpoint.

Changes to _BoundarySnapshotBatchGenerator._process_prompts:

Accept 7th tuple field (prompt_checkpoints) from insert()
Compute effective prompt_checkpoint matching upstream semantics
Replace hardcoded prefill split (1) with prompt_checkpoint in both left-pad and right-pad paths
Add prompt_checkpoint_callback support for upstream parity
Defensive clamp in cache prepare to prevent negative lengths
Backward compatible: When prompt_checkpoints are not supplied (default), prompt_checkpoint computes to 1 and all code paths behave identically to pre-patch.

Tested on M3 Ultra 256GB with mlx_lm 0.31.0 — chat completions, benchmarks, continuous batching all working. Soaking since March 8.

[jundot]

Thanks for this, really appreciate the contribution! I verified the diff against upstream's _process_prompts and everything lines up correctly. The defensive max(0, l - prompt_checkpoint) clamp and the VLM embedding handling during checkpoint processing are nice touches that upstream doesn't need but omlx definitely does.

Merged and working well on my end. Also pushed a follow-up commit (6a9b264) on top of this to finish the rest of the 0.31.1 migration (presence/frequency penalty integration, dependency refs update).