Match mlx-vlm Gemma 4 prefill logits

[samir-lms]

Summary

• Forward mlx-vlm's opt-in logits_to_keep=1 contract through batched VLM chunked prefill, final prefill, and decode
• Preserve exact Gemma 4 generation parity with mlx-vlm 0.6.11 while avoiding full-sequence LM-head projection
• Add focused prefill/decode forwarding coverage

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[samir-lms]

Superseded by #361 after renaming the head branch to samir/gemma4-fix.