[Improvement] save best ckpt during training

[irvingzhang0512]

Fix #460

Add arg save_best_ckpt for eval hook,. If set to True, then save best ckpt in work_dir with name best_{best_socre}_{epoch_id}.pth

• modify eval hook
• add unittest
• add changelog
• fix codecov

[codecov]

Codecov Report

Merging #464 (4500f7f) into master (4df447b) will increase coverage by 0.05%.
The diff coverage is 76.00%.

@@            Coverage Diff             @@
##           master     #464      +/-   ##
==========================================
+ Coverage   84.69%   84.74%   +0.05%     
==========================================
  Files         118      118              
  Lines        8344     8352       +8     
  Branches     1366     1367       +1     
==========================================
+ Hits         7067     7078      +11     
+ Misses        930      927       -3     
  Partials      347      347              

Flag	Coverage Δ
unittests	84.73% <76.00%> (+0.05%)	⬆️

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

Impacted Files	Coverage Δ
mmaction/models/builder.py	55.26% <0.00%> (ø)
mmaction/models/heads/roi_head.py	28.88% <0.00%> (ø)
...maction/models/roi_extractors/single_straight3d.py	81.81% <50.00%> (-3.55%)	⬇️
mmaction/core/evaluation/eval_hooks.py	88.09% <89.47%> (+4.62%)	⬆️

---

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 4df447b...4500f7f. Read the comment docs.

[dreamerlin]

Thanks for your contribution
Lately, we will refactor the EvalHook in the PR #395, which may then be applied into MMCV for other OpenMMLab repo usage:

1. Add by_epoch param to determine perform evaluation by epoch or by iteration.
2. Add automatically assign a compare key when users do not set one.
3. Save checkpoint by creating a symlink as well as record best info in runner hook meta
4. Refactor the code structure

you can modify based on #395

[irvingzhang0512]

Thanks for your contribution
Lately, we will refactor the EvalHook in the PR #395, which may then be applied into MMCV for other OpenMMLab repo usage:

1. Add by_epoch param to determine perform evaluation by epoch or by iteration.
2. Add automatically assign a compare key when users do not set one.
3. Save checkpoint by creating a symlink as well as record best info in runner hook meta
4. Refactor the code structure

you can modify based on #395

Sure, I will close this pr and create a new one tonight.

[dreamerlin]

no hurry, you can make a new pr when #395 is merged

[irvingzhang0512]

no hurry, you can make a new pr when #395 is merged

ok