Invalid page size assumptions in GC

[jkotas]

At very least, there seem to be an issue in GC as well:

Default 4MiB regions have 32KiB mark bitmaps, so adjacent regions share an OS page. 
Commit accounting double-counts that page, while decommit accounting fails to subtract it, causing the assertion.

With temporary  DOTNET_GCRegionSize=800000  (hexadecimal 8MiB), the full normal Checked CoreCLR suite completed

propsed fix:

size_t min_gc_region_size =
    OS_PAGE_SIZE * (mark_word_size / sizeof(uint32_t));

Originally posted by @EgorBo in #133675

[dotnet-policy-service]

Tagging subscribers to this area: @anicka-net, @dotnet/gc
See info in area-owners.md if you want to be subscribed.

[jkotas]

A few more findings from the same 64KiB-page machine, separate from the thunk and GC bitmap-accounting issues above:

1. GC pre-OOM reclamation has another 4KiB assumption. move_aged_regions compares a region's committed size against GC_PAGE_SIZE (4096), whereas its initial commitment is an OS page. Under LLDB, a free region with 65536 committed bytes was skipped by this special pre-OOM reclamation path. This reproduces even with the 8MiB-region workaround. The comparison should use the actual initial commitment.