Fix redirect behavior of datasets.utils.download_url

[pmeier]

Fixes #3560. #3236 introduced redirect resolving in datasets.utils.download_url(). For this requests.get() is used:

vision/torchvision/datasets/utils.py

Lines 64 to 75 in c808d16

	def _get_redirect_url(url: str, max_hops: int = 10) -> str:
	import requests
	for hop in range(max_hops + 1):
	response = requests.get(url)
	if response.url == url or response.url is None:
	return url
	url = response.url
	else:
	raise RecursionError(f"Too many redirects: {max_hops + 1})")

This has two downsides:

1. requests.get() already downloads the data although we only want to check if a redirect happens. Meaning, we download the data twice. This never stood out in our tests, since we download a small amount of data which leads to a minimal increase in testing time. For large datasets, say STL10 STL10 does not download #3560, this is not acceptable. We should switch to a head requests, e.g. requests.head(), that also carries the redirect information without downloading any data.
2. Every dataset that has a the download flag is now indirectly dependent on requests without any mention of this in the documentation.

This PR solves both.

[codecov]

Codecov Report

Merging #3564 (6e10f3a) into master (c808d16) will decrease coverage by 0.04%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #3564      +/-   ##
==========================================
- Coverage   78.84%   78.79%   -0.05%     
==========================================
  Files         105      105              
  Lines        9755     9756       +1     
  Branches     1568     1568              
==========================================
- Hits         7691     7687       -4     
- Misses       1577     1583       +6     
+ Partials      487      486       -1     

Impacted Files	Coverage Δ
torchvision/datasets/utils.py	66.83% <100.00%> (-2.36%)	⬇️

---

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 c808d16...6e10f3a. Read the comment docs.

[fmassa]

Thanks!