Given that we make pretty heavy use of decorators, I believe every layer of decoration adds additional function call overhead whenever the function is called.

We'll have to do some profiling (see #9) to see if this is a real concern, but if it is, it would probably help adoption if we added a compiled backend that allows us to minimize function call overhead by constructing function objects.

I think it might be preferable to do this CPython style where there is a compiled and a pure Python implementation that are kept in sync by the test suite. While that introduces some support burden keeping the two in sync, this is a small enough library (and is intended to be a "lightweight" dependency) that I think it's probably worth it.