~40% improvement over status quo (perf measurements are best-of-5)
### before
```console
$ time flake8 /dev/null
real 0m0.337s
user 0m0.212s
sys 0m0.028s
```
### after
```console
$ time flake8 /dev/null
real 0m0.197s
user 0m0.182s
sys 0m0.012s
```
This includes creating the pool, tearing it down under normal use, and
tearing it down in case of exception.
Doing this makes it harder to leak processes, as for instance was
happening in #410.
Fixes#410
It seems likely that the multiprocessing module on Windows is not
capable of serializing an object with the structure that we have and
preserving the attributes we dynamically set on plugins (like the
FlakesChecker). To avoid issues like this with all plugins (although
we have only found this on Windows with the FlakesChecker), let's try
serializing the Checkers PluginTypeManager to a dictionary so that the
only object that a Queue is really trying to serialize/deserialize is
the FlakesChecker itself.
Related to #179
Due to https://bugs.python.org/issue27649, we cannot continue to
expect multiprocessing to work as we expect and document it on Windows.
As such, we are going to revert back to our previous behaviour of
disabling it across all versions of Python on Windows to provide the
default expected behaviour of Flake8 on that Operating System.
This relies on two things:
1. Properly configuring flake8-import-order to use that style
2. Properly configuring flake8-import-order to know that flake8 is our
application name.
