Pretty much marking a few tests as straight incompatible or marking
certain functions with monkeypatching support.

Also address the lack of 3.8 or higher support in src/black/parsing.py
caused by my original changes.

Finally tweak the compile configuration.

Typing the ast3 / ast27 variables as Any breaks mypyc. More details in
the comment added in this commit.

Many thanks goes to Jelle for helping out, with their insight I was
able to find a workaround.

before .whl was 1.3M and now it is 1.2M

The `platform=linux` was causing mypy (and therefore mypyc) to believe
any `if sys.platform == "win32":` branch would never be taken. That led
to `Reached allegedly unreachable code!` crashes because of safety
checks mypyc adds.

There's not a strong reason for pinning the platform in `mypy.ini` so
with the agreement of Jelle, it's gone now!

I also disabled CliRunner's exception catching feature because while it
does lead to nicer test failures when an exception goes unhandled, it
also removes traceback information which makes debugging a pain.

P.S. You can blame Windows Update for the slowness of this bugfix :p

Unreachable code has either led to mypyc being unable to analyze or
compile Black, or has led to runtime failures (like the one the previous
commit fixed - although mypy wouldn't be able to catch it).

Hardcoding the targets isn't great since we will probably forget to
add paths to the list as new files are created.

Inspired from mypy's setup for mypyc.

- early binding; plus
- reordering of if checks to reduce work / hit the happy case more
  often; plus
- a few more tiny mypyc-specific tweaks

Honestly at this point, I was tired of optimizing and only attempted
basic and safe optimizations. I hope this actually makes a performance
impact :p

Mostly just dataclasses nonsense + more pain from the TOML bare CR
thing ... ugh

diff-shades depends on reformat_many being monkeypatchable so it can
calculate files and black.Mode without copying and pasting a bunch
of parsing, file discovery, and configuration code.

On the other hand, I've been working on blackbench and upcoming version
21.8a1 is now compatible with the driver convert changes.

- I don't actually have to go on a rant about tomli and bare CR it
  turns out.
- Subclassing KeyError isn't actually *that* important for one
  custom exception.