Leverage execution policies inside dbAsIfFlatRegion object traversal

[nikosavola]

Same idea as #2364, but for dbAsIfFlatRegion.

merge_polygons_to buffers the polygons with properties and sorts them by property ID before grouping. That sort now uses std::execution::par where the standard library supports execution policies, and stays a plain std::sort where it doesn't (macOS/libc++, MSVC without TBB, C++11 builds). The guards are __cpp_lib_execution and __has_include(<execution>).

Benchmark

The region_merge_properties case from the db_parallel_benchmarks Google Benchmark suite on the parallelize-rebased branch of my fork runs Region#merged end to end on a flat region of property-tagged boxes, so the timing includes everything around the sort, not just the sort itself.

• 16384 boxes over 8 property groups
• Intel Xeon Gold 6248 @ 2.5 GHz, 16 cores, Ubuntu 24.04, GCC 13.3, -O2, TBB 2021.11
• serial and parallel builds of this branch timed back to back at each CPU count, 5 repetitions with random interleaving per run, medians reported
• CPUs limited with taskset, which also throttles the TBB pool since it sizes itself from the affinity mask

CPUs	serial (ms)	par (ms)	speedup
1	22.816	22.490	1.01x
2	25.463	22.922	1.11x
4	25.572	22.980	1.11x
8	25.282	22.612	1.12x
16	27.185	24.279	1.12x

At 4096 boxes the picture is the same: 1.00x on one core, 1.12x at 16.

The sort is a small part of the whole merge, so the operation gains about 10% and flattens out near 1.12x. Nothing changes single-threaded. I also ran the same benchmark on an Apple M5 (aarch64, Ubuntu container) and the numbers follow the same curve, so this is not x86-specific.