[Feature] Support STARK training

[JingweiZhang12]

No description provided.

[JingweiZhang12]

1. The head of STARK doesn't have functions such as get_targets like siamprn++, and the unit test of stark head is all included in the test_stark_forward, so I don't write repeatedly unit test about stark_head.py.

[codecov]

Codecov Report

Merging #434 (f07b1f2) into master (297c106) will increase coverage by 0.78%.
The diff coverage is 89.42%.

@@            Coverage Diff             @@
##           master     #434      +/-   ##
==========================================
+ Coverage   71.23%   72.01%   +0.78%     
==========================================
  Files         119      121       +2     
  Lines        6748     6954     +206     
  Branches     1294     1326      +32     
==========================================
+ Hits         4807     5008     +201     
  Misses       1546     1546              
- Partials      395      400       +5     

Flag	Coverage Δ
unittests	71.91% <89.42%> (+0.73%)	⬆️

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

Impacted Files	Coverage Δ
mmtrack/models/sot/stark.py	80.85% <80.85%> (ø)
mmtrack/models/track_heads/stark_head.py	91.71% <91.71%> (ø)
mmtrack/models/sot/__init__.py	100.00% <100.00%> (ø)
mmtrack/models/track_heads/__init__.py	100.00% <100.00%> (ø)
mmtrack/models/reid/base_reid.py	78.26% <0.00%> (-8.70%)	⬇️
mmtrack/models/trackers/base_tracker.py	84.73% <0.00%> (-1.53%)	⬇️
mmtrack/datasets/pipelines/transforms.py	88.07% <0.00%> (+0.91%)	⬆️
mmtrack/models/sot/siamrpn.py	81.56% <0.00%> (+1.11%)	⬆️
mmtrack/datasets/sot_train_dataset.py	85.85% <0.00%> (+3.03%)	⬆️
mmtrack/datasets/pipelines/processing.py	80.17% <0.00%> (+5.17%)	⬆️
... and 1 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 297c106...f07b1f2. Read the comment docs.

[JingweiZhang12]

The refactored code of STARK testing is ready, I will create PR of that after this PR is merged.

[GT9505]

In the forward_train of stark, we only need to

1. extract the features of imgs,
2. pass the feature into the head
3. get the training targets from head
4. compute loss
   you can refer to siamrpn

[GT9505]

If there is anything that makes you confused, you can discuss it with me.

[GT9505]

Generally speaking, the model can only call the general API of head, such as head.init_weights, head.loss and so on.
The model should not call the unique attribute of head, such as head.run_cls_head, head.run_reg_head, head.cls_head, head.reg_head and so on.