Lowers tile splitting limit so models with extremely dense mesh portions could still have reasonable performance while for more common cases the performance impact should be minimal.
See T90695 for such case. And please test some medium to big scenes to see if this really hinders performance.
A more clever tile-splitting strategy should be implemented, such as "if couldn't get at least 20% rejection rate in one split, then stop", rather than current "fixed depth" mode.
This patch does _not_ back-port to 2.93.