Latest results with LLM streaming and pre-tool speech

[fanny-riols]

• Enable LLM streaming for cascaded models (2 or 3 parts) and pre-tool speech by default
• Update results on our website with these two new features; refresh leaderboardStats.json with the latest run results:
  • Replaced:
    • Nova 3 + GPT-5.4 + Sonic 3 --> Flux General Multi + GPT-5.4 + Sonic 3.5
    • Gemini 3 Flash + Gemini 3.1 Flash TTS --> Gemini 3.6 Flash + Gemini 3.1 Flash TTS Preview
    • Ink Whisper + Haiku 4.5 + Sonic 3 --> Ink 2 + Haiku 4.5 + Sonic 3.5
    • Parakeet 1.1 + Gemma 31B + Kokoro --> Parakeet 1.1B CTC en-US + Gemma-4-26B + Kokoro
  • Updated "clean" results for:
    • Universal 3.5 Pro + GPT-5.4 + Sonic 3.5
    • Nova 3 + GPT-5.4-mini + Aura 2
  • Removed the "clean" data for the following systems, but kept the perturbation data:
    • Cohere + Gemma-4-26B + Voxtral
    • Whisper Large v3 + Qwen35-27B + Voxtral
• Metric tables (Accuracy/Experience heatmaps) now skip any system whose clean data is empty instead of showing a row of dashes for it
• Rebuilt the docs/ website artifact with the above changes