ENH: give user control over selection of spatial index frame in sjoin

As discussed in #1421, it could be advantageous to give users some control over which frame (between left_df and right_df) is used as the spatial index backend, as this can have a huge impact on sjoin performance.

One proposal was a index keyword that could have values left, right or None (https://github.com/geopandas/geopandas/pull/1421#issuecomment-631342047, https://github.com/geopandas/geopandas/pull/1421#issuecomment-631578974).

ccing @ljwolf, @martinfleis feel free to assign this to me

[bnaul]

In the documentation vein: I agree that a recipe for more advanced patterns is a good idea, but maybe it also makes sense to clarify in the docstring that (if I’m reading correctly) right_df.sindex is used by default unless the predicate is within? I think a simple hint could be enough to help most people avoid accidentally sjoining in a much less performant direction.

[FenderJazz]

I haven’t really got my head around the modular sjoin, but options 2 and 3 (in your list, @adriangb) seem quite complicated for an average user, who just wants to provide a performance hint, without learning about sjoin internals and writing code to fit it.

Documentation is good, of course, but I don’t think it’s true that this issue is resolved by documentation alone. Swapping the order of the GDFs in the sjoin call has implications for the interpretation of all the other arguments, and for the structure of the result, and could require the user to do a lot of recoding and testing to successfully map ‘predicate’ and ‘how’, swap ‘lsuffix’ and ‘rsuffix’ etc. and then re-order columns. It seems a pity to impose this on the userbase, rather than write it once within sjoin!