fix: detect and re-download corrupt vision_encoder weights for Qwen-Edit

When running Qwen-Image-Edit series, check if the pre-quantized base
model's vision_encoder/model.safetensors is smaller than 600MB (expected
~650MB). If so, delete the corrupt file and download marker to trigger
a fresh download from ModelScope.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>