[Model] Keep vision encoder weights unquantized to maintain accuracy

[mengshyu]

This PR ensures that vision encoder layers are excluded from quantization, improving accuracy for models with vision components.