[ONNX] I can not convert table forcast to onnx

### Requested feature
- Batch Inferences
- Convert to ONNX

I have some problems when try to convert tf to onnx:
- I understand this inferences code is process each image, is it anyway to get batch inferences
- More over i have some problem in self.attention in decoder when convert it to onnx, when cache is change but shape when reshape is not.

No files, tests, or entry points are named. Start by identifying the table forecasting model's TensorFlow/ONNX conversion entry point and the decoder self.attention path, then reproduce the per-image inference and cache/reshape failure. Done means the requested batch inference behavior and successful ONNX conversion are demonstrated by tests.

Written by the indexing model from the issue text.