Improve get_interface_attr speed through caching

[fahhem]

From #1599:

Inspired by #1184 but only a small step.

To me, it looks like the car/ module does a lot of (potentially unnecessary) work at import time, such as defining dataclasses for documentation of a car. As a result, collecting the tests has to not only import class and function definitions, but also do real work.

In this case, interfaces.get_interface_attr() walks a folder structure ignoring some folders, dynamically imports modules, does some optional dict-merging, and returns. During test collection, it's called 3 times, each time walking the directory structure and importing the same module.

It seems to me that actually closing #1184 would entail a dozen or more of PRs like this: finding a small improvement to the import-time work or moving it either to runtime or maybe even "compilation".

I switched to functools.cache as requested by @sshane. Speed is still very close, might be slightly faster, but definitely in the noise. Sometimes 0.88s, sometimes 0.93s, but on master I get a similar spread. I'm not sure it's worth analyzing the time much further.