Before the algorithm was written to optimise clipping a few
features against thousands of mask features. The revised algorithm
is optimised for clipping thousands of input features against
a few mask features.

Given that this second operation is much more likely, it makes
sense to optimise for this use case.

I've also applied some other optimisations like taking advantage
of spatial indexes on the providers, using prepared geometries
and also only applying an intersection operation if the geometry
isn't wholly contained by the mask geometry.

Benchmarks:

clipping roads layer with 1 million lines against 2 polygons

before: 5 mins 30 seconds
after: 10 seconds

clipping address layer with 5 million points against 2 polygons

before: 50 minutes
after: 30 seconds

(cherry-picked from 71ebdb8)