Compress cached JSON files

[dosisod]

Closes #15731

[dosisod]

At some point we should probably rename the .json cache files to .json.gz since they're gzipped now. The .json extension is sprinkled in a lot of different places though, so some refactoring would need to be done to fully complete this. Should this be done in this PR, or is it ok to leave it for another PR? Thanks!

[hauntsaninja]

Let's do that in this PR!

[dosisod]

@hauntsaninja it would appear that some of the mypy tests rely on manually creating JSON files:

mypy/test-data/unit/check-incremental.test

Lines 3739 to 3742 in b327557

	-- This is a heinous hack, but we simulate having a invalid cache by clobbering
	-- the proto deps file with something with mtime mismatches.
	[file ../.mypy_cache/3.8/@deps.meta.json.2]
	{"snapshot": {"__main__": "a7c958b001a45bd6a2a320f4e53c4c16", "a": "d41d8cd98f00b204e9800998ecf8427e", "b": "d41d8cd98f00b204e9800998ecf8427e", "builtins": "c532c89da517a4b779bcf7a964478d67"}, "deps_meta": {"@root": {"path": "@root.deps.json", "mtime": 0}, "__main__": {"path": "__main__.deps.json", "mtime": 0}, "a": {"path": "a.deps.json", "mtime": 0}, "b": {"path": "b.deps.json", "mtime": 0}, "builtins": {"path": "builtins.deps.json", "mtime": 0}}}

Now that gzipped files are expected, what would be the best way to get around this? One of the things I considered was adding a gzip_file directive to the test data runner that's similar to file, except it gzips the data before writing it to disk. Does that seem like a good route, or do you see a better option?

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

graphql-core (https://github.com/graphql-python/graphql-core): typechecking got 1.05x slower (295.0s -> 311.0s)
(Performance measurements are based on a single noisy sample)

[JukkaL]

Since this breaks some workflows that parse or compress the mypy cache files, I wonder if this should be treated as a breaking change, which would benefit from a feature flag to enable/disable compression. In the first release with the feature we'd have the flag off by default, and we could turn it on by default in the next feature release. See our policy for backward incompatible changes for more information.