Segregate inter level hostility

[kevingranade]

Summary

SUMMARY: Performance "Reduce inter-faction hostility overhead with a inter-level visibility cache."

Purpose of change

Execution time when passing time when near a large number of monsters hostile to each other is slow, and profiling says it's dominated by inter-faction vision checks.

Describe the solution

One of the dominant use cases where this happens is inside or near a lab complex since there are many monsters hostile to each other separated by walls. This use case is ameniable to special handling by seperating the maps of creatures based on levels, and only checking for visibility between monsters on levels that have some kind of visibility between them.
I added a bool per map level cache entry that tracks whether the level has a contiguous floor. This is calculated essentially for free in build_floor_cache.
An accessor walks these entries and tabulates which levels the origin level has visibility to as a bitset.
I split the monster faction map by level, within the entry for each faction, there is a sparse map of levels, each of which contains a data structure holding the actual weak pointers to the monsters.
When scanning the faction map for enemies, the code can now exit early on entries on levels not visible to the current monster.

Describe alternatives you've considered

This still consumes a significant amount of processing time, though it is no longer dominant, so further optimization is warranted.

Testing

A timing test of waiting outside a lab for three hours tool 52 seconds before this change, and 40.75 seconds after this change. Profiling with perf also indicated that this particular codepath is doing drastically less work. The new cache creation code is not appearing in profiling.

[Aivean]

An idea for the future, we can use the new no_floor_gaps cache to speed up sunlight_cache further (avoid iterating over the floor if we know there are no gaps).

UPD: after a bit of thinking, that won't actually speed up anything, as the fully_inside -> fully_outside transition doesn't happen directly, unless the current map is completely flat. What can help there is to check submap::is_uniform.

[kevingranade]

Updated to opportunistically shortcut out of checking for NPCs and player as well.

[ZhilkinSerg]

It is not safe to compare two floats like that.

[kevingranade]

@alvean we could add the opposite cache value of all_open or similar built the same way and use it to determine the first layer with obstacles more cheaply.

@ZhilkinSerg the only time I expect that code to trigger is when the relative positions of monsters are rotations of each other, i.e. 1,7 vs 7,1, and in that case I do expect them to trigger reliably. I'm not against adding some kind of epsilon here to make it trigger a bit more, but that requires a bit more thought and I don't want to mess with it in this pr.

[Aivean]

@alvean we could add the opposite cache value of all_open or similar built the same way and use it to determine the first layer with obstacles more cheaply.

Good idea, however I'm going to try checking submap::is_uniform first, it might work well enough.