Improve partition box key hashing

[yuqian19870725-creator]

Summary

• replace the linear implicit size_t conversion with a dedicated KeyHash
• hash each coordinate independently to reduce unordered-map bucket collisions
• add regression coverage for coordinates that previously collided and for 16³ distinct box lookups

Validation

• MSVC v143 standalone compile/runtime check passed using the updated Partition.h
• 16³ distribution check improved from 106 distinct old hashes to 4,096 distinct new hashes
• the full CMake configure could not complete because the project triplet requires the unavailable Visual Studio v141 toolset

[sverrirsccp]

Hey there - thanks a lot for your contribution!

I work at Fenris Creations and wanted to let you know that we've acknowledged your PR. We'll take a proper look at it as soon as we can and really appreciate you taking the time to contribute.

[ccpgames-carbon]

All committers have signed the CLA.

[hrafn]

I agree that the current hash function here does not spread things out as much as we would like, but am not entirely convinced of this new approach, as it seems a bit over-elaborate, and adds a test that doesn't seem to add any value.

[yuqian19870725-creator]

@hrafn Thanks again for the detailed feedback. I added benchmark results and rationale to the hash-combiner thread, and would appreciate a fresh review of the current revision when you have a chance. If you still prefer a simpler approach or want the collision test changed/removed, I am happy to update the PR accordingly.

[hrafn]

Ok, you make a convincing argument for your hashing approach, but the tests need cleaning up.

Remove the RetainsEveryDistinctBoxKey test, and I'd be happy to merge this in.

[yuqian19870725-creator]

@hrafn Done — I removed the RetainsEveryDistinctBoxKey test as requested. The focused regression test remains. Thanks for the review; this should be ready to merge.