Add faster validation to improve training speed

[MiXaiLL76]

Describe the feature

Faster implementation of the COCOeval function written in C++

Motivation

I often work with the mmdet project and use datasets in COCO format. There are a large number of objects in my datasets (more than 3000 for 1 photo). I would like to get validation after each epoch, but at the same time not delay training. The standard COCOeval algorithm with such a number of objects is slow, but there is a faster implementation, which I cleaned out of dependencies (torch / detectron2) and use in my work.
I am ready to open a PR and transfer the developments to the project.

Related resources

The original implementation of the library is in detectron2
Also, at some point, christofferedlund started working on clearing the library of facebook dependencies, but abandoned the project without putting the source codes on github.
I found the source codes on the Internet and continued his work. faster_coco_eval

Additional context

I benchmarked the validation on the original coco val dataset and presented the results in the project repository.

Visualization of testing comparison.ipynb available in comparison
Tested with yolo3 model (bbox eval) and yoloact model (segm eval)

Type	COCOeval	COCOeval_faster	Profit
bbox	22.854 sec.	8.714 sec.	more than 2x
segm	35.356 sec.	18.403 sec.	2x

[RangiLyu]

We will opensource a repo for evaluation recently, we may add this feature to that repo.

[MiXaiLL76]

We will opensource a repo for evaluation recently, we may add this feature to that repo.

It's a good idea

[MiXaiLL76]

#120