Release GenStructures' cached column state after the postpass

[Zaldaryon]

Summary

Investigates #270 (BlockSchematicStructure objects never collected) by pulling and reading the actual profiler screenshots attached to the issue, then fixes the mechanism they point at.

Every retention chain in the screenshots ends the same way: BlockSchematicStructure/BlockSchematicPartial arrays reachable through WorldGenStructuresConfig.LoadedSchematicsCache via GenStructures, or through GenDungeons via a similar path, terminating at what looks like per-thread storage rather than a plain object reference. One screenshot names the exact field: ServerEventManager.stratumCachedWorldSavedInvocations.

That pointed at Stratum's own concurrency patch to GenStructures.cs (added to let TerrainFeatures generate multiple columns at once). It adds [ThreadStatic] fields (stratumStateOwner, stratumHeightmap, and a few siblings) so the postpass can reuse a chunk's already-computed heightmap and climate/forest values instead of recomputing them. The problem: that cached state is only ever overwritten when the same worker thread happens to process another chunk's main pass. Once worldgen quiets down, whatever a thread last touched (the owning GenStructures instance and its loaded schematic cache included) stays reachable from that thread for as long as the thread itself lives. Vanilla's own unpatched code doesn't have this property, since the equivalent state sits in plain instance fields tied to the GenStructures instance's own lifetime, not to a specific thread.

Fix

Release stratumStateOwner and stratumHeightmap at the end of the postpass, the point at which a column's structure generation is actually finished. Both fields get set fresh by the main pass regardless of which branch runs, so clearing them right after the postpass reads them is always safe: the next chunk on that thread recomputes from scratch rather than reusing anything.

Caveat

This is a real, demonstrable gap from vanilla's GC behavior, but a single heap snapshot can't distinguish "leaked forever" from "a worldgen thread that's mid-column right now." I can't reproduce a long-running server session with active generation followed by an idle period to directly confirm the counts stop growing. @tehtelev , since you filed this with the profiler already set up: if you get a chance to rerun the same capture after a while on this branch, that would close the loop for real.

Type

• Bug fix

Checklist

• scripts/extract-patches.sh ran clean. One file touched: GenStructures.cs.patch.
• dotnet build VintageStory.slnx -c Release is green (0 errors, 0 new warnings).
• Every vanilla edit has a // Stratum marker. The new lines carry one explaining why they're there.
• No vanilla source committed.
• Tested on a real server start, not just compilation (make smoke: PASS, reached RunGame).

Related issues

Refs #270

[tehtelev]

It looks like everything is fine now.