separate caching logic from download for PhotoTour dataset

[pmeier]

This separates the caching logic from the download() method. Ultimately, the caching should be removed similar to #3420.

Since we have no tests and the combination of downloading and caching makes it hard to test we separate it first (this PR), add tests in a follow-up PR and only then remove the caching.

[codecov]

Codecov Report

Merging #3448 (41865eb) into master (b266c2f) will decrease coverage by 0.14%.
The diff coverage is 20.00%.

@@            Coverage Diff             @@
##           master    #3448      +/-   ##
==========================================
- Coverage   76.13%   75.98%   -0.15%     
==========================================
  Files         105      105              
  Lines        9703     9701       -2     
  Branches     1556     1556              
==========================================
- Hits         7387     7371      -16     
- Misses       1836     1844       +8     
- Partials      480      486       +6     

Impacted Files	Coverage Δ
torchvision/datasets/phototour.py	25.24% <20.00%> (-0.01%)	⬇️
torchvision/datasets/cifar.py	90.00% <0.00%> (-5.00%)	⬇️
torchvision/datasets/caltech.py	79.31% <0.00%> (-4.60%)	⬇️
torchvision/datasets/voc.py	92.30% <0.00%> (-2.20%)	⬇️
torchvision/datasets/vision.py	68.96% <0.00%> (-1.73%)	⬇️
torchvision/transforms/_transforms_video.py	88.88% <0.00%> (-0.35%)	⬇️
torchvision/models/mobilenetv2.py	86.95% <0.00%> (-0.28%)	⬇️
torchvision/transforms/_functional_video.py	95.23% <0.00%> (-0.22%)	⬇️

---

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 b266c2f...fc9c0ae. Read the comment docs.

[fmassa]

Thanks!