Supporting class-specific boxes in `CombinedNonMaxSuppression` op

[masahi]

Currently, CombinedNonMaxSuppression converter does not support class-specific boxes, i.e. q == num_class case in https://www.tensorflow.org/api_docs/python/tf/image/combined_non_max_suppression.

tensorflow-onnx/tf2onnx/onnx_opset/tensor.py

Lines 1873 to 1874 in 01ec092

	utils.make_sure(share_boxes_across_classes,
	"CombinedNonMaxSuppression only currently implemented for boxes shared across classes.")

Apparently, "official" TF FasterRCNN / MaskRCNN models uses such variant of combined nms. So we cannot convert them to ONNX right now.
https://github.com/tensorflow/models/blob/238922e98dd0e8254b5c0921b241a1f5a151782f/official/vision/detection/modeling/architecture/heads.py#L360-L361

I've hit this issue when trying to convert QAT-ed MaskRCNN model from https://github.com/openvinotoolkit/nncf/tree/develop/examples/tensorflow/segmentation. The model is attached for repro.
https://drive.google.com/file/d/1vaNmyKH0C0n5PGlo2oZ2wggME1Blg5sf/view?usp=sharing

I wonder if it is possible to encode this variant using a combination of ONNX NMS + other ops. For example, by making boxes into the shape (batch, num_boxes * num_class, 4) and doing "one hot" encoding for scores. Ideally we could change the ONNX spec to allow this variant, since the implementation is actually trivial (just need to change indexing into boxes).

@TomWildenhain-Microsoft