[KernelBench] Add XeGPU support in kernel bench tool

[tkarna]

• Adds a simplified xegpu matmul schedule that takes tile size parameters as kwargs. This is easier to hook up with the yaml desciptors.
• Adds xegpu target support to kernel-bench tool. Currently enabled only in the torch.compile code path.
• test_kernel_bench.py example:
  • Add --pipeline arg to optionally override the schedule yaml file look up.
  • Adds schedules/xegpu/matmul/bf16.yaml example schedule for a ~4k matmul. Compatible with the existing schedule lookup scheme.

Examples (executes, verifies, okayish performance):

python test-kernel-bench.py --kernel level1/1_ --dtype bf16 --target xegpu --torch-compile --benchmark --nruns 500 --nwarmup 500
python test-kernel-bench.py --kernel level1/2_ --dtype bf16 --target xegpu --torch-compile --benchmark --nruns 500 --nwarmup 500

[tkarna]

Possible improvements:

1. Generalize benchmark yaml descriptors. CPU and GPU targets might require different metadata (e.g. problem size). Flop count calculation should be symbolic to take into account the used problem size.
2. Generalize schedule yaml descriptors. The MLP schedule, for example takes parameters as a list of dicts, one for each layer. At the moment we cannot express this. The schedule should be separated from the parameters (e.g. use the same schedule for many benchmarks, just change the params).
3. Current TargetInfo arch/feature sounds CPU specific, generalize (?).
4. Support tile size tuning. This is required to get good performance on XeGPU. We need a way to obtain the KB payload IR and run it with the matmul tuner. Either at runtime or as a separate step and (in both cases) cache the params on disk.

[rengolin]

#207 propagates the target flags, will conflict with this PR

[rengolin]

one comment, otherwise, lgtm

[tkarna]

this should be sufficient as the torch.bfloat16 case never matches

[rengolin]

It does in #211, but that can be done in that PR, not here.

[adam-smnk]

I'll double check if both are needed, too many changes recently 😅