Mostly by providing a "numpy shim" for CircuitPython,
try to make the numpy tests run on all three systems.
(a "scipy shim" might also be useful?)
However, there are test failures. Is it worth working
through them and getting this to a point where it could
be included?
