Remove the bwraps checks from _site.py and put them in platform.linux
which is the only place where they are run.

This allows the removal of a double level of caching, making
reasoning about tests easier

This removes the `_instance` on the platform object that we use
for caching and not recreating the object everytime at the start
of every test.

This is to ensure our tests share the least amount of state.

The performance penalty is from 2 to 5% accross the whole test suite.
The readings were done 5 times for each before and after the change
and on the same computer.

We need to supress the brwap version caching for it to work.

This caching leaks for every test but removing it every time seems
too prohibitive, given that, normally, the bwrap version would not
change.

On a fedora 29 laptop, the performance degradation was around 5-10%
over the whole test suite if used globally.