Add retry logic for vision model (handles 500/504 timeouts)

Fix: replace non-existent en-US-DavisNeural with ChristopherNeural

Try Qwen2.5-VL-72B-Instruct for vision

Switch vision model to Qwen/Qwen2.5-VL-7B-Instruct (supported by inference providers)

Fix vision model: switch to Llama-3.2-11B-Vision + resize images to 1024px max

Add Chair/moderator who introduces the topic and speakers

Add animated robot avatars synced to speaker turns

Use CohereLabs/aya-vision-32b - confirmed available on HF inference

Switch to Qwen models - available on HF inference without gating

Fix asyncio: use new_event_loop instead of nest_asyncio

Remove explicit provider - let auto-router select with valid HF token

Fix InferenceClient: explicitly use hf-inference provider

Initial commit: GeoTalk AI Geoscience Debate Podcast