VolatilePtr load not always paired with proper store instruction

We've observed crashes in Roslyn's language server on macOS ARM64 while accessing the thread pointer in `DeadlockAwareLock::TryBeginEnterLock`https://github.com/dotnet/dotnet/blob/a1b9d991e4596823841f4f9d840e72ac369e3a7e/src/runtime/src/coreclr/vm/threads.cpp#L7417

@davidwrighton has noticed in some stacks we are loading the Volatile field that was stored without a proper barrield in other paths.