Still requires the more conservative __index__ here rather than a possibly
truncating __int__ because this is used in a context where floating point
values should probably be treated specially.

This fixes Github issue #2752.