Incorrect Cross Attention Values from Generate Function of ORTModelForVision2Seq

[KateWinslet]

Check ONNX export compatibility, verify attention settings, and ensure ONNX Runtime supports generate and attention layers for image-to-text tasks.