NativeAOT-LLVM: Do not use mmap/munmap

[yowl]

This PR adds an emscripten friendly VirtualReserve/Release implementation. Emscripten does not fully implement mmap/munmap in particular it does not zero the memory in mmap for the MAP_ANONYMOUS flag. This continues from dotnet/corert#8330.

When trying to merge the latest into NativeAOT-LLVM random memory crashes, usually inside the GC, have started to appear, so now seems a good time to ask for feedback on this change. This change does stop those crashes appearing so far at least.

Emscripten reports the page size as 0x4000, whereas CommonMacros.h has

#elif defined(HOST_WASM)

#define DATA_ALIGNMENT  4
#ifndef OS_PAGE_SIZE
#define OS_PAGE_SIZE    0x4
#endif

I don't know what the implications are there, if anything maybe just performance as posix_memalign does not refer to the page size.

[jkotas]

This is really only used for argument validation. The GC should not be ever calling these APIs with invalid arguments.

I think you can omit this structure, and instead directly call posix_memalign / free / memset to implement reserve / release / decommit.

[yowl]

Thanks, works with without that.

[jkotas]

LGTM. Thanks!