feat(vlm): add vision feature caching for LFM2-VL

[MrFelt0385]

feat(vlm): add vision feature caching for LFM2-VL

Summary

LFM2-VL currently skips the opportunistic vision feature cache (the code
says // LFM2-VL runs every image in one preparation pass; skip the opportunistic vision cache for this first integration.). This PR implements
it, giving a 2.12× speedup on repeated images — the main win for
multi-turn conversations and continuous screen analysis.

Changes

• src/vision/lfm2_vl.rs — add get_input_embeddings_with_cache,
  mirroring the existing Qwen2.5-VL pattern:
  • Cache key: SHA-256 of the request image bytes + soft-token budget
    (image_hash_from_bytes_with_soft_tokens on the server side)
  • HIT: reuse the concatenated post-connector features, skip the vision
    tower + connector entirely
  • MISS: compute features, eval + deep-copy (decouple from the
    deferred graph), store snapshot in the cache
• src/multimodal/vlm_runtime.rs — wire the Lfm2Vl runtime to the
  single-feature cache (ModelVisionCaches::single), deriving the key from
  the concatenated pixel tensor (same approach as the Qwen path)

Benchmarks

Measured on a Mac mini M1, LFM2.5-VL-3B-MLX-4bit, same image across
requests (zone_hud.png, 120×70):

Request	Time	Notes
1st (MISS)	1158 ms	vision tower encodes the image
2nd (HIT, same image)	547 ms	2.12× faster
3rd (HIT, same image, different text)	550 ms	2.10× faster

Finesse preserved: HIT responses are byte-identical to MISS responses
(vision features are deterministic; only the tower output is cached).

Notes

• The connector output shape depends on the image's native patch count,
  which is folded into the key by the soft-token budget — two resolutions of
  the same image cannot collide.
• Cache lifetime is tied to the model registry entry (clear_all on unload),
  same as the existing Qwen/Gemma/Granite caches.

[cla-assistant]

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 have signed the CLA already but the status is still pending? Let us recheck it.}