feat: implement sequential prefill fallback for multimodal runner and add support for Qwen3.5 with vision encoder

[barhanc]

Description

• Implements sequential prefill fallback for multimodal runner required to support Qwen3.5 with vision capabilities.
• Adds symbols for Qwen3.5 VL model
• Adds model picker component to multimodal screen in LLM example app, similar to LLM screen

NOTE: Support for Qwen3.5 VL model is experimental right now since due to architectural constraints (namely the GatedDeltaNet implementation) model requires sequential prefill fallback which results in very slow prefill. Additionally the model output right now is not very good (repetition, etc.) which probably can be fixed by applying #1099 .

The best action for now is probably to wait until the ExecuTorch team adds better support for Qwen3.5 architecture.

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

• Run example app and test new VLM models

Screenshots

Related issues

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

Qwen3.5 VLM model was exported using https://github.com/barhanc/executorch/tree/rne-v1.2.0-export-qwen-3.5 .

[msluszniak]

For now you can rebase and:

1. Resolve conflicts
2. Test if merged feat(llm): min_p and repetition_penalty sampling, per-model defaults, letterbox vision #1099 actually fixed repetition. This will help us detect if the problem is model base or was caused by the incorrect configuration.

[barhanc]

Based on my limited testing it looks like the proper configuration actually somewhat helped. However, the models themselves aren't particularly good, especially compared to something like LFM2.5. And as of right now, they aren't really usable due to slow prefill.

ExecuTorch team merged the PR with custom op for GatedDeltaNet, but they reverted it later (pytorch/executorch#19178), so I guess we still have to wait.