reevaluate numba minimum, usage patterns in spa.py

conda search -c main numba | grep py36 shows that numba 0.36.1 is the oldest package compatible with python 3.6 and numpy 1.12. The numba git history says 0.36.1 was released on Dec 7, 2017. (0.17.0 was released Feb 3, 2015.) I’ll add that to the asv conf.

I’d be fine with also adding a minimum numba requirement to setup.py, but I think that should be done in combination with changes to the import logic in spa.py. In particular, I don’t like the numpy fallback - that led to hard to track down errors when developing this. It would also be worth reviewing modern numba best practices. If we’re adding a minimum numba requirement to setup.py then we should probably be testing against it too. More than I want to tackle in this PR.

_Originally posted by @wholmgren in https://github.com/pvlib/pvlib-python/pull/1059#discussion_r488775560_

[wholmgren]

I’m fine with giving up call-by-call numba control. We could also give up PVLIB_USE_NUMBA and instead point users to NUMBA_DISABLE_JIT. Situations in which a user wants numba for one library but not another are probably pretty rare, so I’m skeptical that it’s worth our time to code around that.

[wholmgren]

I was thinking we’d rely on different ci configurations for total coverage. The coverage within a single configuration would be incomplete. So I don’t think we’d need any special import or environment machinery.