Add Qwen 3.5 and Gemma 4 models to benchmark suite

[vimeto]

New models to benchmark

For the MobiHoc 2026 paper revision, we should include newer model families to strengthen the generalizability of our findings.

Models to add

Qwen 3.5 family:

• Qwen 3.5 (small variants, ~1-4B) — successor to Qwen 3, likely improved tool calling
• Check mlx-community for INT4 variants
• SGLang should support these with qwen parser

Gemma 4 family:

• Gemma 4 (small variants) — successor to Gemma 3n
• May resolve the SGLang/vLLM compatibility issues we had with Gemma 3n
• Check if standard attention (no Conv3d) improves runtime support

What to do

1. Check HuggingFace for available model variants and sizes
2. Add model configs to MLX_MODEL_MAP and MODELS lists
3. Create optimized prompts in optimized_prompts.py
4. Validate tool calling works on both MLX and SGLang
5. Run full benchmark sweep

Context

• Current models: Qwen 3 (4B, 0.6B), Llama 3.2 3B, DeepSeek R1 1.5B, Gemma 3n E2B
• Newer models would test if our structural findings (capability threshold, prefill dominance, inverse efficiency) hold across model generations
• Reviewer A: "the paper feels like a snapshot in time" — newer models address this