[Feature] Add AvoidOOM to avoid OOM

[BIGWangYuDong]

First, trying to change torch.mm to torch.einsum to avoid OOM:
before change, mAP: 0.331

inter_matrix = torch.mm(flatten_masks, flatten_masks.transpose(1, 0))

after change, mAP: 0.331

inter_matrix = torch.einsum('ik, kj -> ij', flatten_masks,  flatten_masks.transpose(1, 0))

But found it cannot save GPU memory.

	torch.mm	torch.einsum
process time (s)	4.69856	4.5713
GPU memory (MiB)	807	807

To avoid OOM, we add a class, which can try to convert inputs to FP16 and CPU if got a PyTorch's CUDA Out of Memory error.
It will do the following steps:

1. first retry after calling torch.cuda.empty_cache().
2. If that still fails, it will then retry by converting inputs to FP16.
3. If that still fails try to convert inputs to CPUs. In this case, it expects the function to dispatch to the CPU implementation.

TODO:

• Add docs in FAQ

Close: #6908

[jbwang1997]

If we can only set keep_dtype to make users decide whether the outputs are still the same as the input data type or are converted to the avoid OOM data type (fp16, cpu).

[codecov]

Codecov Report

Merging #7434 (929ea0d) into dev (151a803) will decrease coverage by 0.58%.
The diff coverage is 41.02%.

@@            Coverage Diff             @@
##              dev    #7434      +/-   ##
==========================================
- Coverage   65.09%   64.50%   -0.59%     
==========================================
  Files         357      360       +3     
  Lines       28852    29233     +381     
  Branches     4891     4954      +63     
==========================================
+ Hits        18782    18858      +76     
- Misses       9061     9370     +309     
+ Partials     1009     1005       -4     

Flag	Coverage Δ
unittests	64.49% <41.02%> (-0.61%)	⬇️

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

Impacted Files	Coverage Δ
mmdet/utils/memory.py	40.25% <40.25%> (ø)
mmdet/utils/__init__.py	100.00% <100.00%> (ø)
mmdet/models/detectors/__init__.py	100.00% <0.00%> (ø)
mmdet/models/dense_heads/__init__.py	100.00% <0.00%> (ø)
mmdet/models/dense_heads/solo_head.py	65.18% <0.00%> (ø)
mmdet/models/dense_heads/solov2_head.py	9.83% <0.00%> (ø)
mmdet/models/detectors/solov2.py	83.33% <0.00%> (ø)
mmdet/core/bbox/assigners/max_iou_assigner.py	73.68% <0.00%> (+1.31%)	⬆️
mmdet/models/roi_heads/test_mixins.py	52.85% <0.00%> (+2.14%)	⬆️
mmdet/models/dense_heads/dense_test_mixins.py	43.20% <0.00%> (+2.46%)	⬆️
... 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 151a803...929ea0d. Read the comment docs.

[BIGWangYuDong]

Update the logic in AvoidOOM, which defaults to return source type and device without any interface. This makes the codes look simpler.