[RFC] Add support for joint transformations in VisionDataset

[fmassa]

This is a WIP PR that adds support for jointly transforming both the input and the target.

This has been requested a lot, specially for tasks such as semantic segmentation and object detection.

This is a first draft, which implements the approach that has been proposed in #230

Thanks to the VisionDataset from @pmeier , handling backwards-compatibility will be fairly straightforward without having to write a lot of code.

Feedback is welcome. I'll start with a few datasets which are generally used for detection / segmentation, and we will iterate on the design.

[codecov-io]

Codecov Report

Merging #872 into master will increase coverage by 0.38%.
The diff coverage is 26.53%.

@@            Coverage Diff             @@
##           master     #872      +/-   ##
==========================================
+ Coverage    54.4%   54.79%   +0.38%     
==========================================
  Files          36       36              
  Lines        3327     3338      +11     
  Branches      549      550       +1     
==========================================
+ Hits         1810     1829      +19     
+ Misses       1380     1373       -7     
+ Partials      137      136       -1

Impacted Files	Coverage Δ
torchvision/datasets/voc.py	22.1% <0%> (+1.71%)	⬆️
torchvision/datasets/sbd.py	31.25% <0%> (+0.48%)	⬆️
torchvision/datasets/coco.py	29.26% <25%> (+4.77%)	⬆️
torchvision/datasets/vision.py	40.35% <34.37%> (-2.99%)	⬇️
torchvision/datasets/imagenet.py	21.55% <0%> (ø)	⬆️
torchvision/transforms/transforms.py	83.37% <0%> (+0.68%)	⬆️
torchvision/transforms/functional.py	70.47% <0%> (+1.26%)	⬆️
torchvision/datasets/utils.py	36.84% <0%> (+1.73%)	⬆️

---

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 90091b5...f0fde4c. Read the comment docs.

[fmassa]

@vfdev-5 FYI I've renamed xy_transform to transforms to make it consistent with the other transforms. Given that this was added after the 0.2.2 release, I think it's fine to break backwards-compatibility.