[ci-test] disable lse outlining

[EgorBo]

It seems that lock()'s bottleneck (no contention, ThinLock) is CompareExchange. Since it's a Thin Lock we effectively know there is no thread contention. It seems like in that case we might want to avoid LSE outlining since LSE is more about high-contention scenarios.

this is a no-merge PR, just to run a benchmark.

[EgorBo]

@EgorBot -azure_cobalt100 -azure_ampere -aws_graviton3 -aws_graviton4 -gcp_axion

using System.Runtime.CompilerServices;
using BenchmarkDotNet.Attributes;

public class Benchmark
{
    static readonly object SyncRoot = new();

    [Benchmark]
    public void Object()
    {
        lock (SyncRoot)
            DoWork();
    }

    [MethodImpl(MethodImplOptions.NoInlining)]
    private static void DoWork() { }
}

[EgorBo]

looks like it does improve perf on all arm cpus: EgorBot/runtime-utils#536

[jkotas]

LSE is more about high-contention scenarios

LSE is also about less expensive memory barriers. The perf results may be different once you do a bunch of memory reads and writes on both inside and outside the lock to mimic more realistic scenario, and the more expensive memory barrier in non-LSE variant has actual work to do.