S2LatLngRect::GetCentroid test failures

A few years ago, the comparison threshold for rect centroid splitting was increased from 1e-15 to 2e-15.

https://github.com/google/s2geometry/blob/ac448e2e939863dfefdeb6500ea74fda92d8187c/src/s2/s2latlng_rect_test.cc#L646-L649

Now I notice that the "full latitude range" tests sometimes fail with 1e-15 and need ~4/3 e-15.

https://github.com/google/s2geometry/blob/ac448e2e939863dfefdeb6500ea74fda92d8187c/src/s2/s2latlng_rect_test.cc#L684

@ericveach Do you remember if these were supposed to be rigorous, or just a small value that seemed to work?