API: auto-detect when to use numba

[jbrockmendel]

ATM groupby reductions (and window i think?) have keywords that allow for opting-in to using numba, along with engine_kwargs that control options passed to numba.

I think users would be better served by automatically detecting cases where numba is available/useful and automatically using it. Simpler API, better performance for the large subset of users who don't read docs.

[lithomas1]

Hm, not sure if the numba implementation of operations in pandas is quite there yet.

I've been having a look at the numba kernels lately (mainly in groupby) and the 2 major limitations that I've found are:

1. Numba kernels are hard-coded to return float64
   (This is a bit annoying to deal with, but I should have a PR for this, late this week or early next week).
2. No support for our nullable masked arrays.

There's also parts of pandas where the numba support is half done (e.g. Groupby.agg doesn't support list/dict of functions).

https://github.com/pandas-dev/pandas/issues?q=is%3Aopen+is%3Aissue+label%3Anumba is a good place to look for other issues.