[Benchmarks] Pin benchmarks to small set of cores

[PatKamin]

For better results stability, pin benchmark binaries to four cores with the maximum available frequency.

[PatKamin]

Test run on the PVC perf machine: https://github.com/intel/llvm/actions/runs/18679758239/job/53258464661

[lslusarczyk]

some suggestions

[lukaszstolarczuk]

LGTM

[PatKamin]

@intel/dpcpp-devops-reviewers, is adding numactl package to docker images ok?

[PatKamin]

@intel/dpcpp-devops-reviewers, please review. This change will help stabilize benchmark results on the new BMG machine.

Note that benchmark job that actually uses numactl installed passed: https://github.com/intel/llvm/actions/runs/19104203407/job/55173686283.
The benchmark job that runs automatically for this PR doesn't have numactl package installed yet, so its result can be neglected.

As to the e2e test failure, it's unrelated.

[PatKamin]

@intel/dpcpp-devops-reviewers, a friendly ping. This patch should also fix failing benchmarks on BMG machine: https://github.com/intel/llvm/actions/runs/19328176408/job/55285113356#step:19:2263

[lukaszstolarczuk]

thx, LGTM

[sarnex]

@PatKamin Is this ready for merge? Is the benchmark CI fail unrelated?

[lslusarczyk]

@PatKamin Is this ready for merge? Is the benchmark CI fail unrelated?

Yes. As @PatKamin informed me, it is ready to merge and benchmarks fail will be fixed by dockers rebuilding this night.
Please merge.