Integrate C-IL for instance segmentation

[harimkang]

Summary

This PR includes :

• Integrate class-incremental learning for instance segmentation task (recipe)
• Add model template {EfficientNetB2B-maskRCNN & Resnet50-maskRCNN}
• Add ignore mode for instance-segmentation
• Modify the code of the infer function to be the same as OTE detection inference_task
• related OTE PR: [MPA] Class-incremental Learning for instance-segmentation training_extensions#1142

[harimkang]

@goodsong81 @JihwanEom @jaegukhyun @chuneuny-emily @supersoob
Hi, I share this because I have a problem to think about together.

The current default value for task_adapt_op is 'MERGE'. It is done in most tasks, and it is not a problem in C-IL setting. The problem is that sometimes org_model_classes are imported when using 'coco pretrained weights' or other pretrained weights in a 'Supervised Learning setting'. (ex, 80 classes for coco)

To solve this problem, we can fix the code as above (155-157 lines) with 'REPACE' as default. This is the problem that appears in (the instance-segmentation ResNet pretrained weight) and detection (Yolox pretrained weight) so far.

If you have any concerns or ideas, please share them. Also, I wonder if this happens in other tasks.

[goodsong81]

No objection for 'REPLACE' default. In SC/OTE the data_classes is the final target classes anyway.

[goodsong81]

LGTM.