Fine-grained: Treat files changed only if md5 changes

[JukkaL]

Previously the daemon could do a lot of extra work if you switched
to another git branch and immediately back to the original branch,
since many file timestamps would change. Now this isn't sufficient
to consider a file as changed.

Also add a new cached file system abstraction (that is currently only
used in dmypy) to avoid redundant file system operations and to make
file system state easier to reason about. The idea is that we cache
output of previous file system operations during a single increment,
for both consistency and performance. My plan is to eventually use it
about everywhere.

This should also make it slightly easier to switch to listening to
file system events instead of stat()ing everything, but that's not
a priority yet.

Fixes #4499.

[JukkaL]

Travis is failing on master as well -- the failures seem unrelated to this.

[msullivan]

This all seems basically reasonable.

My only concern is that now we have two separate mechanisms for tracking the hashes of files to see if they have changed, but they are at very different places in the abstraction stack, so it isn't too concerning.

[JukkaL]

My only concern is that now we have two separate mechanisms for tracking the hashes of files to see if they have changed, but they are at very different places in the abstraction stack, so it isn't too concerning.

This is a concern, but this shouldn't be significantly worse now than it was previously. In the near/medium term I want to get the regular build using the new abstraction as well. This should give a minor perf boost and also make things a bit cleaner in general.

[MakonnenMak]

@JukkaL is this item from the docs still relevant with this change?

https://mypy.readthedocs.io/en/stable/additional_features.html#remote-cache

If you use the mypy daemon, you may want to restart the daemon each time after the merge base or local branch has changed to avoid processing a potentially large number of changes in an incremental build, as this can be much slower than downloading cache data and restarting the daemon.