Reimplement AutoAssign

[jshilong]

I reimplemented autoassign.
I got a similar accuracy of 40.4 vs 40.5 in origin implementation, and the performance is unstable with 1x setting and may fluctuate by about 0.3 mAP. Results between 40.3 and 40.6 are acceptable.

[codecov]

Codecov Report

Merging #4295 (9ef65a2) into master (bca2766) will increase coverage by 0.43%.
The diff coverage is 94.32%.

@@            Coverage Diff             @@
##           master    #4295      +/-   ##
==========================================
+ Coverage   65.59%   66.02%   +0.43%     
==========================================
  Files         255      259       +4     
  Lines       20036    20272     +236     
  Branches     3407     3448      +41     
==========================================
+ Hits        13142    13385     +243     
+ Misses       6188     6172      -16     
- Partials      706      715       +9     

Flag	Coverage Δ
unittests	65.99% <94.32%> (+0.44%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmdet/models/dense_heads/fcos_head.py	70.08% <ø> (+13.39%)	⬆️
mmdet/models/dense_heads/autoassign_head.py	93.92% <93.92%> (ø)
mmdet/models/dense_heads/__init__.py	100.00% <100.00%> (ø)
mmdet/models/detectors/__init__.py	100.00% <100.00%> (ø)
mmdet/models/detectors/autoassign.py	100.00% <100.00%> (ø)
mmdet/models/necks/fpn.py	95.40% <100.00%> (+0.16%)	⬆️
mmdet/models/detectors/cornernet.py	94.87% <0.00%> (-5.13%)	⬇️
mmdet/apis/inference.py	38.18% <0.00%> (-5.12%)	⬇️
mmdet/models/dense_heads/corner_head.py	73.72% <0.00%> (-1.88%)	⬇️
mmdet/datasets/coco.py	46.25% <0.00%> (-0.06%)	⬇️
... and 10 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update bca2766...9ef65a2. Read the comment docs.

[Johnson-Wang]

General comments:
The overall code structure is fine to me despite too many hard-coded hyperparameters. Is it possible to specify loss_pos and loss_neg in AutoAssignHead, so that others may be free to change to other loss styles?

[ZwwWayne]

1. Please resolve the conflicts and comments.
2. After resolving the comments, @jshilong can benchmark models for release