This removes the assertions that checked plugin devs didn't do anything
overly bad in favour of using mypy to ensure appropriate type-checking.
It also moves the ProgressCallback typing information to constants in
order to avoid circular imports (since constants doesn't import
anything).
Since we stack using Mac layers as well as Linux layers, both sets of
banner caches are required. As the cache builders are separate at the
moment, we previously ran banner searches without any banners. This was
bad. It inflated memory hugely and killed the buildbot.
We now bomb out on stacking a Linux or Mac layer unless we have banners
against which to match.
Relented on the strict import of direct objects/classes for the typing
module only. Typing module components can be directly imported because
it makes the code really painful to read and write otherwise.
This is still in-line with the python style guide adopted from Google at
http://google.github.io/styleguide/pyguide.html section 2.2.
Unfortunately in tests it turned out re was faster than a python
implementation of Wu-Manber. The Wu-Manber code has been left (because
why not) but it's not really advantageous over the re builtin since
it's compiled in C.