ARM64: avoid memory barrier in InlinedMemmoveGCRefsHelper if dest is not on heap

[EgorBo]

Should speed up on-stack struct copies

using System.Runtime.CompilerServices;
using System.Text.Json;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Running;

BenchmarkSwitcher.FromAssembly(typeof(Bench).Assembly).Run(args);

public class Bench
{
    [Benchmark]
    public void DestNotInHeap()
    {
        Utf8JsonReader dst = default;
        Utf8JsonReader src = default;
        DoCopy(ref dst, src);
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    public void DoCopy(ref Utf8JsonReader dst, Utf8JsonReader src) => dst = src;
}

Linux-arm64 (Ampere):

| Method        | Mean     | Error    | StdDev   |
|-------------- |---------:|---------:|---------:|
| DestNotInHeap | 15.39 ns | 0.014 ns | 0.013 ns | (Main)
| DestNotInHeap | 11.46 ns | 0.010 ns | 0.008 ns | (PR)

Fixes #102044

[dotnet-policy-service]

Tagging subscribers to this area: @mangod9
See info in area-owners.md if you want to be subscribed.

[EgorBo]

PS: Utf8JsonReader (and, afair, some other STJ structs) is quite big (192 bytes) and is byref-like (ref struct), so, presumably, it should benefit a lot from #9512

[EgorBo]

@jkotas I guess this is ready for review then?

[jkotas]

This method seems to be structured to allow inlining of BulkMoveWithWriteBarrier at every callsite.

I think it would be better to keep these checks inside the non-inlined portion of the helpers. The empty buffer or same buffer are rare cases. We do not need to optimize for them by inlining these checks at every callsite.

[jkotas]

LGTM. Thank you!