Add Kaggle 80-class hard-mode benchmark for Qwen3-VL-4B q4

Replace approx-sign tildes with U+2248 to avoid GFM strikethrough misparse

Read KHANA_API_KEY from env var instead of hardcoding it

Add full VLM benchmark suite and results: model comparison, prompt engineering, ZeroGPU deploy

Add RAW_PROMPT: mode to analyze() for direct benchmark comparison

Surface exception details in analyze() response for debugging

Add torchvision (required by Qwen3-VL's video sub-processor)

Fix invalid gradio sdk_version, update README for Qwen3-VL-4B

Upgrade KhanaVision Space to Qwen3-VL-4B-Instruct

Initial KhanaVision inference Space