Globe: symbol and coveringTiles optimizations

[kubapelc]

This PR ports symbol rendering optimizations from non-globe main branch to current globe-based main and adds some other globe-specific optimizations on top:

• Added a fast path for symbol projection for mercator projection, comparable to MapLibre 4.7 in performance - this just fixes a performance regression that was present in the globe branch.
• Improved collision box placement performance by not calling isTilePositionOccluded (the information was already computed beforehand).
• Improved globe's coveringTiles performance by caching tile AABBs instead of recomputing them every time.
• Fixed globe's coveringTiles incorrectly returning hundreds of hidden tiles as visible if the camera is sufficiently pitched.
• Improved globe's coveringTiles tile visibility accuracy, reducing the number of returned tiles and thus improving performance.
• Added a SymbolCollisionBoxGlobe benchmark for symbol placement under globe projection. This is the same as SymbolCollisionBox, just with globe projection used instead of mercator.
• Fixes bad line scaling near the poles
• Fixed globe loading tiles at unnecessarily high zoom level when pitched

Below are benchmark results for mercator projection. Here "main" essentially represents globe branch without optimizations ported over, thus it is relatively slow.

These next two figures are the new globe placement benchmark, first on the current main branch, then on this PR branch.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 88.02395% with 20 lines in your changes missing coverage. Please review.

Project coverage is 88.08%. Comparing base (835abdb) to head (a6a7f3d).
Report is 16 commits behind head on main.

Files with missing lines	Patch %	Lines
src/geo/projection/globe_transform.ts	50.00%	6 Missing and 12 partials ⚠️
src/geo/projection/mercator_transform.ts	93.75%	0 Missing and 1 partial ⚠️
src/symbol/collision_index.ts	98.79%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4778      +/-   ##
==========================================
+ Coverage   87.99%   88.08%   +0.09%     
==========================================
  Files         265      265              
  Lines       37591    37659      +68     
  Branches     2337     2343       +6     
==========================================
+ Hits        33078    33172      +94     
+ Misses       3482     3456      -26     
  Partials     1031     1031              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[HarelM]

I've added a few comments. thanks for this!
There is a PR that also looks at covering tiles here, it would be great to get your input there:

• Fix level of detail at high pitch #4750

(it's currently closed due to globe branch deletion, it should be opened soon).

[HarelM]

New PR is here:

• Fix level of detail at high pitch #4779

@NathanMOlson - can you check that the covering tiles optimization and improvements done here are incorporated in your PR - I would like to separate responsibilities: have Nathan's PR address covering tiles stuff and have this PR focus on symbol optimizations.

[kubapelc]

@HarelM @NathanMOlson thank you for your feedback! I will address it and improve this PR in a week or two, sadly I need to focus on other things in the meantime. I think this PR will not create merge conflicts with the terrain tile LOD PR, since that changes different parts of the covering tiles logic.

[NathanMOlson]

can you check that the covering tiles optimization and improvements done here are incorporated in your PR

The two optimizations for covering tiles in this PR are

1. Caching of tile AABBs. This needs changes to address unbounded memory growth.
2. Early return if tileBelowHorizon(). I'm not convinced this is actually an improvement.

I don't think these changes are ready to be incorporated in my PR.

[HarelM]

I would advise to remove covering tiles related logic from this PR and optionally add it to @NathanMOlson's PR if applicable.

[kubapelc]

I have reverted most of coveringTiles changes, with one exception at line 84:

return Math.min(centerDist, cameraDist) * 2 <= radiusOfMaxLvlLodInTiles * tileSize; // Multiply distance by 2, because the subdivided tiles would be half the size

This is a simple bugfix, before this change globe displayed unnecessarily many tiles at high zoom levels. This alone improves performance a lot when camera is pitched.

I will try to implement a better tile culling by using general bounding frustums instead of AABBs. Frustums can approximate the actual tile shape much more tightly, thus leading to more discarded tiles, thus improving performance. A frustum-frustum visibility test should not be much slower than frustum-AABB test, but I will measure the times and create a benchmark. I will also implement a tile bounding volume cache that does not leak memory. I will do that as a separate PR.

[HarelM]

Added a few more comments. Otherwise looks good enough to be merged. Thanks!

[HarelM]

Can you update the branch from main to resolve conflicts?
I believe other comments were addressed.

[kubapelc]

Should be fixed now, I also fixed the changelog wording and the failing render test (reverted the expected image to its state before this PR).