Add XeGPU matrix multiplication benchmark

[tkarna]

This issue outlines the required steps to add an XeGPU matrix multiplication benchmark.

1. mlir-gen should be a Python module.
   • Move to python dir tree as a module, e.g., python/lighthouse/payload_generator/payload_generator.py
   • Add command line interface, e.g., python/lighthouse/payload_generator/cli.py
   • Map to an executable script, e.g., mlir-gen in pyproject.toml:

     [project.scripts]
     mlir-gen = "lighthouse.python.ingress.payload_generator:cli"

2. Generic infrastructure to define and execute workloads
   • A Workload object to hold the payload IR and fixed metadata (e.g. problem size).
   • Provides methods for getting payload IR, schedule IR, input arguments for calling payload, (optionally) correctness verification, etc.
   • A generic execution_engine wrapper that can execute a Workload and can, e.g., run it in a timer loop for benchmarking.
3. XeGPU matmul Workload and benchmarking tools
   • Currently supports only matmul + elementwise post ops.
   • Location: python/lighthouse/benchmark/xegpu/matmul.py (?)
   • Generates payload, defines lowering schedule, compiles, executes, measures performance.
   • Uses existing functionality: workload, payload generator, execution toos, etc.
   • Exposed as another CLI command, e.g., benchmark_xegpu_matmul
4. Installation mechanism for XeGPU support:
   • Compile LLVM with LevelZero runtime and necessary flags.
   • Hook up LLVM Python bindings with Lighthouse.
     • Simply set PYTHONPATH="$LLVM_INSTALL_PATT/python_packages/mlir_core/"
   • Provide an easy-to-use install mechanism.
     • Build instructions in README(?).
     • Or a generic build script, invoked on install(?).
       • mlir-python-bindings Python package is not installed in this case(?).
     • When executing kernels, raise a descriptive error if Xe GPU device/drivers are missing.

Examples of the benchmark command line tool usage:

$ benchmark_xegpu_matmul
sizes=4096,4096,4096 dt=f16,f32 wg-tile=256,256 sg-tile=32,32 ... time(ms): 1.76 GFLOPS: 78072

$ benchmark_xegpu_matmul --sizes 4096 2048 1024 --wg-tile-size 256 256 ...
...

$ benchmark_xegpu_matmul --dump-kernel {intial,tiled,vectorized,bufferized,xegpu-wg,...}
<prints payload IR at this stage of lowering>

$ benchmark_xegpu_matmul --dump-kernel xegpu-wg --dump-schedule
<also dumps transform schdule used to lower to this level>