Update benchmarks with M4 Pro results and add Mamba vocab mismatch note

- Transformer (Qwen2.5-0.5B): 40 tok/s, 28% acceptance
- Mamba results marked experimental due to vocabulary mismatch