fix: pass modelSupportVision to convertToModelMessages in summary generator and generateText (#562)

Without this flag, images are sent as raw image data instead of OCR text
to models that don't support vision, wasting tokens and causing errors.