Optimize read_video_timestamps for some formats

[fmassa]

I feel this is fragile as it currently stands.
It would be great to identify which types of codec the packet pts and the frame pts perfectly align, but I couldn't find a precise information yet.

This seemed to work for videos in Kinetics. It greatly speeds up getting the timestamps, which is important for large video datasets.

Out of a better solution (with more guarantees from the codec sides), I'm pushing this as is for now.

cc @stephenyan1231

[codecov-io]

Codecov Report

Merging #1168 into master will increase coverage by <.01%.
The diff coverage is 80%.

@@            Coverage Diff             @@
##           master    #1168      +/-   ##
==========================================
+ Coverage   65.91%   65.91%   +<.01%     
==========================================
  Files          79       79              
  Lines        5809     5818       +9     
  Branches      882      886       +4     
==========================================
+ Hits         3829     3835       +6     
- Misses       1711     1713       +2     
- Partials      269      270       +1

Impacted Files	Coverage Δ
torchvision/io/video.py	71.42% <80%> (+0.48%)	⬆️
torchvision/transforms/transforms.py	80.94% <0%> (-0.2%)	⬇️

---

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 8102158...b93bd5c. Read the comment docs.