Add better docs for FasterRCNN, MaskRCNN and KeypointRCNN

[fmassa]

This also adds a bugfix in the default value for the MaskRCNNPredictor, which was discovered when adding a different backbone.
This doesn't change the default behavior nor the default models, but is simpler when one wants to specify a custom backbone.

[codecov-io]

Codecov Report

Merging #943 into master will decrease coverage by 0.07%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #943      +/-   ##
==========================================
- Coverage   60.26%   60.19%   -0.08%     
==========================================
  Files          63       63              
  Lines        5001     5002       +1     
  Branches      745      745              
==========================================
- Hits         3014     3011       -3     
- Misses       1784     1787       +3     
- Partials      203      204       +1

Impacted Files	Coverage Δ
torchvision/models/detection/faster_rcnn.py	74.39% <ø> (ø)	⬆️
torchvision/models/detection/keypoint_rcnn.py	81.81% <ø> (ø)	⬆️
torchvision/models/detection/mask_rcnn.py	81.35% <100%> (+0.32%)	⬆️
torchvision/transforms/functional.py	69.9% <0%> (-1.22%)	⬇️

---

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 05bc255...801c5db. Read the comment docs.