Add Gemma 12b Unified Support

[neilmehta24]

Changes:

• Delete the legacy "VisionModelKit" and "Vision Add On", now that everything is being sent through the batched vision model kit
• Add support for gemma 4 12 unified. I added some patches to the model to enable APC, as mlx-vlm disables chunked prefill if there is an image in the prompt. Some code is added to ensure that we don't restore the cache in the middle of an image.
• Add support for vision feature caching. In v1.8.5, while the KV cache would not need to be re-computed, the image embeddings did need to be recalculated. This creates a substantial speedup for follow-up image requests for large models and large image data. Hook into the mlx-vlm implementation for this.

Fixes:

• Fix MLX thread shutdown bug with a workaround: install_mlx_compile_cache_cleanup_for_thread. This bug is already fixed upstream, so we can remove this workaround after the next MLX release
• Add model compatibility fix for LFM2 VL
• Add patches to Qwen 3.5 to restore generation speed to v1.8.5, as some of the upstream MTP additions caused a slowdown

[will-lms]

Was thinking about this path wrt caching. My understanding is that we'll route here if a prompt has images and to Qwen3NextAttention if prompt does not.

Have we tested the case where a vision prompt comes in, restores a cache that was previously generated using the text-only route, and then generates? It is not obvious to me that it would work as expected, but hopefully should be close enough either way. Just wondering if we tested such a branch (or if I'm misunderstanding the caching).

[neilmehta24]

This works as intended, there's no explicit test case for this though

[will-lms]

Reviewed with agentic help. Overall am aligned with decisions.