Add support for LLaVA (multimodal) ExecuTorch models

[JasBogans]

Problem description

Background
React Native ExecuTorch today provides first‑class hooks for pure‑vision (e.g. useClassification, useObjectDetection, useOCR) and pure‑language (useLLM) models—but there’s no built‑in way to load and run a single multimodal (“vision‑language”) checkpoint such as LLaVA-1.5 or BLIP‑2.

Why Multimodal?

• LLaVA, BLIP‑2, Flamingo and similar models can ingest images + text prompts and produce visually grounded responses.
• Enabling this on‑device in React Native would unlock powerful offline scenarios (e.g. photo Q&A, visual assistants) with privacy and low latency.

Proposed solution

A new hook, e.g.

const { result, generate, isLoading } = useVisionLLM({
  modelSource: require('../assets/llava-1.5.pte'),
  tokenizerSource: require('../assets/llava-tokenizer.json'),
});

Alternative solutions

Or extend useLLM to accept an image tensor:

const { result, isReady } = useLLM({
  modelSource: require('…/llava.pte'),
  imageInput: myImageTensor,
  prompt: "Describe what you see",
});

Benefits to React Native ExecuTorch

• One unified multimodal hook removes boilerplate for separately running useImageEmbeddings + useLLM.
• Enables richer on‑device AI experiences (visual QA, instruction following, AR captions) in pure React Native apps.

Additional context

No response