Enable hierarchical classification in MPA

[JihwanEom]

This PR includes:

Summary

• Disable torchreid jumping in MPA when the task runs hierarchical classification
• Add hierarchical model templates for MPA
• Add hierarchical training capacity in MPA
• Revised classification inference for stable operation
• Integrated separated load_annotation method to one method in mpa_cls_dataset.py
• Set drop_last=True when run hierarchical classification
• Support hierarchical classification CLI
• Relocation classification sample datasets in one folder

[harimkang]

I left a few comments.
Additionally,

1. When I looked at the added sample data, I could look at the data sets that already exist, such as 'data/car_tree_bug'. Is there a reason to add an image directly? I think you just need to specify the path in the annotation file, what do you think? (The capacity may increase due to overlapping image data.)
2. There seems to be a conflict in the submodule of the MPA. Please check.

[harimkang]

Some code looks complicated. It would be good to divide it into functions, but it would be good to consider refactoring with SAMClassifier later PR.

[jaegukhyun]

LGTM

[jaegukhyun]

LGTM

[goodsong81]

LGTM.
But please note that we need refactoring in future PRs.
There are too many / long if-else blocks that should be extracted as method or subclassed.