Refactor _scene_cull() to use two stages

[myaaaaaaaaa]

Splits it into a filter stage (where >99% of runtime is spent) and a binning stage. A followup patch converts the binning stage (which is very prone to race conditions, see #71705 ) to single-threaded so that just the filter stage is multithreaded, greatly simplifying the culling code and removing the need to keep track of a local InstanceCullResult in each thread.

This also opens the doors for potentially adding an option to turn off filtering altogether (resulting in no culling), making it easier to detect culling regressions.

I've also observed a slight increase in culling performance (going from 1 -> 2 threads results in an average of 1.3x speedup rather than 1.2x), although obviously YMMV.

Required for converting _scene_cull() to use parallel foreach(): #78016
Required for occlusion soft-culling: #76297

[RandomShaper]

Warrants a good review from the rendering team, but I can't spot anything fishy.

My only comment is that I'd favor (uint32_t)1 over 1ULL. More verbose, but clearer IMO.