fix: unpack prompt_checkpoints in _chunked_next for mlx-lm >= 0.31.0

[Thump604]

Summary

mlx-lm 0.31.0 added prompt_checkpoints as the 7th element in the BatchGenerator.insert() tuple. _chunked_next only unpacked 6 elements, causing ValueError: too many values to unpack (expected 6) when prefix cache triggers the chunked prefill path.

One-line fix: add _prompt_checkpoints to the zip(*batch_prompts) unpack.

Same class of bug as the _process_prompts tuple fix — different code path (_chunked_next bypasses _process_prompts entirely when prefix cache mid-prefill is active).

Reproduction

1. Start server with --enable-prefix-cache and mlx-lm >= 0.31.0
2. Send any request with a prompt long enough to trigger chunked prefill
3. Server crashes with ValueError: too many values to unpack (expected 6)

Test plan

• Start server with --enable-prefix-cache --continuous-batching on mlx-lm 0.31.x
• Send a request with >8K token prompt
• Verify no ValueError on chunked prefill path
• Verify prefix cache hit on subsequent request with same prefix

Fixes #178

[Thump604]

CI green. Small fix — unpack prompt_checkpoints tuple in _chunked_next for mlx-lm >= 0.31.0. Fixes #178.

[Thump604]

Closing in favor of #221, which fixes the same tuple unpack crash but also preserves prompt checkpoint semantics through the chunked prefill lifecycle. Our fix here just discards the checkpoint data silently.

Three other PRs (#183, #156, #221) address the same issue. #221 is the most complete.