Position-sensitive ROI Pool/Align

[sampepose]

This adds PS ROI pool and align. I also fixed a couple bugs with existing ROI pool tests.

[codecov-io]

Codecov Report

Merging #1259 into master will decrease coverage by 0.74%.
The diff coverage is 39.58%.

@@            Coverage Diff             @@
##           master    #1259      +/-   ##
==========================================
- Coverage   65.89%   65.14%   -0.75%     
==========================================
  Files          75       76       +1     
  Lines        5785     5876      +91     
  Branches      885      886       +1     
==========================================
+ Hits         3812     3828      +16     
- Misses       1708     1783      +75     
  Partials      265      265

Impacted Files	Coverage Δ
torchvision/ops/poolers.py	94.02% <100%> (ø)	⬆️
torchvision/ops/__init__.py	100% <100%> (ø)	⬆️
torchvision/ops/ps_roi_align.py	34.04% <34.04%> (ø)
torchvision/ops/ps_roi_pool.py	37.2% <37.2%> (ø)
torchvision/ops/roi_pool.py	67.44% <0%> (-6.64%)	⬇️
torchvision/ops/roi_align.py	65.95% <0%> (-5.48%)	⬇️
torchvision/extension.py	38.09% <0%> (-2.82%)	⬇️
torchvision/utils.py	65.38% <0%> (-0.66%)	⬇️
torchvision/transforms/transforms.py	80.94% <0%> (-0.63%)	⬇️
torchvision/models/shufflenetv2.py	85.54% <0%> (-0.18%)	⬇️
... and 6 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 a129b6b...aa1f5e6. Read the comment docs.

[LukasBommes]

I'm already using the layer and it works fine!

[fmassa]

I will get this reviewed and merged into torchvision soon.
But before we need to have CI for windows and OSX working, so that we can make sure that everything is working fine on the OS that we support, see #1298

[fmassa]

Thanks for the PR!

torchvision requires the ops to be implemented on both CPU and GPU (with a potential fallback to CPU in some particular cases, but CPU implementation is a must).

So this PR can't be merged in it's current state unfortunately.

[fmassa]

torchvision follows similar requirements as PyTorch with respect to the implementation of the ops: it requires the ops to be implemented for both CPU and GPU.

[LukasBommes]

@sampepose Are you working on this? Or should I give it a try?

[LukasBommes]

I'm currently working on the CPU implementation as I rely on this layer for another project.

[fmassa]

This has been subsumed by #1410. Thanks for the original version @sampepose!