[ingress][torch] Add benchmark functionality to torch.compile backend

[tkarna]

• JITFunction now has a benchmark(*args, nwarmup, nruns) function that can be used to time the kernel.
• MLIRBackend.jit_function holds a reference to the latest compiled JITFunction.
• Updated xegpu torch_matmul.py example to use this benchmark method.

Basic usage:

model = ... # create torch model
backend = gpu_backend(...)
model.compile(backend=backend)
outputs = model(input_a, input_b)
# accessing backend.jit_function raises an error if a model has not been compiled yet
times = backend.jit_function.benchmark(input_a, input_b, nwarmup=500, nruns=500)
avg_time = times.mean() * 1e6  # microseconds

[tkarna]

The torch_matmul.py exercises a simple torch model that does not have any parameters (e.g., layer weights). If the model has parameters, those must be prepended to the jit_fn.benchmark call arguments:

times = backend.jit_function.benchmark(*model.parameters(), *inputs, nruns=nruns, nwarmup=nwarmup)