Add fp16 support in action recognition

[dreamerlin]

• Use fp16 utils from mmcv
• Add fp16 config for TSN
• Add self.fp16_enable attribute and @auto_fp16 tag in BaseRecognizer to support fp16 training.
• Ran fp16 setting and normal setting. they got the same accuracy

[codecov]

Codecov Report

Merging #25 into master will increase coverage by 1.58%.
The diff coverage is 93.00%.

@@            Coverage Diff             @@
##           master      #25      +/-   ##
==========================================
+ Coverage   84.67%   86.25%   +1.58%     
==========================================
  Files          73       72       -1     
  Lines        3889     4016     +127     
  Branches      632      623       -9     
==========================================
+ Hits         3293     3464     +171     
+ Misses        490      448      -42     
+ Partials      106      104       -2     

Flag	Coverage Δ
#unittests	86.25% <93.00%> (+1.58%)	⬆️

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

Impacted Files	Coverage Δ
mmaction/core/__init__.py	100.00% <ø> (ø)
mmaction/datasets/pipelines/__init__.py	100.00% <ø> (ø)
mmaction/models/backbones/resnet.py	94.27% <0.00%> (ø)
mmaction/models/backbones/resnet3d.py	89.69% <ø> (ø)
mmaction/models/backbones/resnet_tsm.py	95.08% <ø> (ø)
mmaction/models/heads/base.py	85.71% <ø> (ø)
mmaction/models/recognizers/base.py	60.81% <37.50%> (-0.62%)	⬇️
mmaction/apis/inference.py	81.81% <66.66%> (+15.78%)	⬆️
mmaction/models/losses/ohem_hinge_loss.py	75.75% <75.75%> (ø)
mmaction/datasets/pipelines/formating.py	93.85% <85.71%> (+0.34%)	⬆️
... and 16 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 41e2732...44d748f. Read the comment docs.