TSan: IncrementalGcStartSweep's test hook reads a block a GC worker is reclaiming #436

The TSan job failed on main at 5a9b560 with two data race reports, both in test/regression/IncrementalGcStartSweep:
https://github.com/openfl/libhxcpp/actions/runs/42460790472/job/128669432620 (ubuntu-26.04, image 20261101.186)

The PR runs built on the same commit passed TSan (#425, #429, #434), so it doesn't happen on every run.

Read of size 4 at 0x726c00330618 by main thread:
  #0 __hxcpp_gc_incremental_test(int, hx::Object*)
  #1 Main_obj::main()

Previous write of size 8 at 0x726c00330618 by thread T1:
  #0 memset
  #2 void BlockDataInfo::reclaim<false>(BlockDataStats*)
  #3 GlobalAllocator::ThreadLoop(int)

Location is heap block of size 1576 allocated by main thread:
  #1 GlobalAllocator::AllocMoreBlocks(bool&, bool)

The second report is the same pair the other way round: a write in GlobalAllocator::ThreadLoop to a 1 MB block, after a read of it by __hxcpp_gc_incremental_test on the main thread. T1 is a GC worker that hx::IncrementalDrain started, from an automatic incremental collection that a large allocation set off.

So the test hook on the main thread reads a block's data while a worker is still reclaiming it. The hook is only built with HXCPP_GC_INCREMENTAL_TEST, so this may be the hook reading without waiting for the workers rather than a race a program can hit, but that hasn't been checked. It isn't #430's pattern (LocalAllocator::DoAlloc reading a block AllocMoreBlocks published), which the TSan job no longer reports on ubuntu-26.04.

Next

Check whether __hxcpp_gc_incremental_test waits for the reclaim jobs before it reads block data, and whether anything outside the hook reads blocks the same way.