Remove FLATTEN_MAPS thread local, store it in ZalsaLocal instead

[Veykril]

No description provided.

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	0ff0ffd
🔍 Latest deploy log	https://app.netlify.com/projects/salsa-rs/deploys/6a2913627c9ea90008b884ad

[codspeed-hq]

Merging this PR will degrade performance by 4.21%

❌ 1 regressed benchmark
✅ 7 untouched benchmarks
🗄️ 18 archived benchmarks run¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Memory	converge_diverge_nested	12.2 KB	12.7 KB	-4.21%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing Veykril:lukaswirth/push-nltxvtsyqlrr (0ff0ffd) with master (f6325bb)}

Footnotes

1. 18 benchmarks were run, but are now archived. If they were deleted in another branch, consider rebasing to remove them from the report. Instead if they were added back, click here to restore them. ↩

[MichaReiser]

Can you say more what the benefits of this change are? Is it that the memory gets released when dropping the ZalsaLocal instance over retaining it forever?

I'd have to measure this change on ty but I'm a little concerned about a typical multithreaded use case where we create a new rayon task for each file, and each task uses its own Db clone (its own ZalsaLocal)

[Veykril]

I generally favor having less thread locals, given we already have the thread local state stashed away in ZalsaLocal. Generally I wanted to check the effect of the benching numbers here. You do bring up a good point regarding the fact that ZalsaLocal's tend to get dropped and re-created frequently though (sounds like another thing we might want to have a benchmark for in fact).

Gonna put this in draft again until I can give that some more thought.