Fix mono_class_setup_interfaces race-condition on ARM64

[nulldatamap]

Summary

On ARM64 the weak memory ordering model makes the mono_class_setup_interfaces code racy, causing segfaults.
The fix is to use a load-acquire when reading klass->interfaces_inited.

Fixes UUM-144819.

The Problem

I have a detailed explanation of the problem in the comment included with the fix.

Testing

I've tested that solution works against a repro and I had a general sanity assert to ensure the bitfield offset was correct (see commit 6b098b5)

• Should this pull request have release notes?
  • Yes
  • No
• Do these changes need to be back ported?
  • Yes
  • No
• Do these changes need to be upstreamed to mono/mono or dotnet/runtime repositories?
  • Yes
  • No

Reviewers: please consider these questions as well! ❤️

Release notes

Fixed UUM-144819 @nulldatamap:
Mono: Fix race condition during interface initialization on ARM64.

Backports
6000.0, 6000.3, 6000.5, 6000.6

[cla-assistant-unity]

All committers have signed the CLA.

[nulldatamap]

I ran the Windows ARM64 build with b15853c and it succeeded: https://yamato.ds.unity3d.com/job/71688177. Normally it would crash on the assert because the build runs the editor.
But just to be 100% sure, I'll also trying to do the same thing with the assert, but with a bogus byte/bit offset, just to confirm the check works.

[nulldatamap]

Alright I have proof of the old offsets failing on WinARM64: https://yamato.ds.unity3d.com/job/71759109
And proof of them succeeding with the ones your Claude instance suggested: https://yamato.ds.unity3d.com/job/71818332
So I'll remove the debug check and land with the fixed offsets.