Update #2151 to post .NET 10

[lisandroct]

Description of Changes

Re-implements #2151 on top of latest master and .NET 10 support.

Reuse C# runtime buffers when consuming BytesSource values from module FFI exports instead of allocating fresh byte arrays for each call.

It's reasonable and the benchmarks results are enticing. These changes obviously make sense as long as Wasm module execution stays single-threaded. If future module execution becomes concurrent (or we re-enter these exports recursively), these buffers become unsafe.

I ran the benchmarks two times (with high variability in some cases):

Benchmark	Master	This Branch	Delta
large args 64KiB	48.648 us / 47.275 us	40.893 us / 42.611 us	-15.94% / -9.87%
circles load=10	34.706 ms / 34.631 ms	33.170 ms / 33.325 ms	-4.43% / -3.77%
circles load=100	34.624 ms / 34.500 ms	33.296 ms / 33.163 ms	-3.84% / -3.88%
ia_loop load=10	10.271 ms / 10.362 ms	9.946 ms / 10.056 ms	-3.16% / -2.96%
ia_loop load=100	26.831 ms / 27.533 ms	25.543 ms / 25.627 ms	-4.80% / -6.92%
filter string index	48.699 us / 43.832 us	49.177 us / 46.586 us	+0.98% / +6.28%
filter u64 index	21.535 us / 21.278 us	25.251 us / 21.471 us	+17.26% / +0.91%
insert u32/u64/u64 btree	305.730 us / 247.585 us	293.044 us / 247.641 us	-4.15% / +0.02%
iterate u32/u64/str	49.282 us / 42.789 us	41.700 us / 40.554 us	-15.39% / -5.22%
iterate u32/u64/u64	21.775 us / 18.167 us	15.486 us / 15.167 us	-28.88% / -16.51%

The results are good (negative numbers mean faster execution) in all benchmarks except filter string index, filter u64 index and maybe insert u32/u64/u64 btree. They don't match the original branch results. I assume this comes down to a difference in hardware and .NET SDK version (I ran the benchmarks on .NET 10 with NativeAOT-LLVM).

These changes are probably worth applying.

API and ABI breaking changes

No breaking changes.

Expected complexity level and risk

1. I don't see a lot of risk as long as all tests pass, because it's not adding new functionality.

Testing

• Tests are passing.