Remove caching from MNIST and variants

[pmeier]

Closes #3555

MNIST is one of the oldest datasets in torchvision. Back than we opted to cache the data in a custom binary for speed reasons. The caching happens in the download() method:

vision/torchvision/datasets/mnist.py

Lines 149 to 164 in 22c548b

	print('Processing...')
	training_set = (
	read_image_file(os.path.join(self.raw_folder, 'train-images-idx3-ubyte')),
	read_label_file(os.path.join(self.raw_folder, 'train-labels-idx1-ubyte'))
	)
	test_set = (
	read_image_file(os.path.join(self.raw_folder, 't10k-images-idx3-ubyte')),
	read_label_file(os.path.join(self.raw_folder, 't10k-labels-idx1-ubyte'))
	)
	with open(os.path.join(self.processed_folder, self.training_file), 'wb') as f:
	torch.save(training_set, f)
	with open(os.path.join(self.processed_folder, self.test_file), 'wb') as f:
	torch.save(test_set, f)
	print('Done!')

This makes them harder to test: we need to run this method in each test if we provide a mock of the original binaries, but the new testcases from #3402 were designed to handle the download flag in a special way.

Since the speed difference is not significant anymore, we can safely remove the caching.

---

1. We currently do not have tests for [EQ]MNIST. I've tested everything locally, but please review them extra carefully.
2. Working on this module was really cumbersome. Although the datasets share so much functionality, it is hard to maintain them. I'll add tests for all MNIST variants next and afterwards refactor this module.

[codecov]

Codecov Report

Merging #3420 (8cb1fb8) into master (22c548b) will increase coverage by 0.02%.
The diff coverage is 45.76%.

@@            Coverage Diff             @@
##           master    #3420      +/-   ##
==========================================
+ Coverage   75.13%   75.16%   +0.02%     
==========================================
  Files         105      105              
  Lines        9722     9735      +13     
  Branches     1563     1567       +4     
==========================================
+ Hits         7305     7317      +12     
  Misses       1930     1930              
- Partials      487      488       +1     

Impacted Files	Coverage Δ
torchvision/datasets/mnist.py	57.32% <45.76%> (+2.07%)	⬆️

---

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 22c548b...b0475de. Read the comment docs.

[fmassa]

Looks great, thanks!

I've a couple of comments, let me know what you think

[pmeier]

Blocked by #3443

[fmassa]

Looks great, thanks a ton!