Introduce optional profiling option for SPIRVRunner

[anmyachev]

Closes #2565

Usage: ./build/SPIRVRunner tensor_2 --enable-profiling

Output example:

...
Tensor output: [8192, 8192], Float (268435456 bytes)
Kernel execution time: 4.81872 ms
...

Profiling ref: https://github.com/intel/pti-gpu/blob/master/chapters/device_activity_tracing/DPCXX.md#how-to-use
Profiling ref with Kineto: pytorch/pytorch#52936 (comment)

Example with Kineto

  #include <torch/csrc/autograd/profiler_kineto.h>
  using namespace torch::autograd::profiler;
  ProfilerConfig cfg{
    ProfilerState::KINETO,
      false,
      false,
      false,
      false,
      false
  };
  std::set<torch::autograd::profiler::ActivityType> activities{torch::autograd::profiler::ActivityType::CPU, torch::autograd::profiler::ActivityType::XPU};
  prepareProfiler(cfg, activities);
  enableProfiler(cfg, activities);

  sycl_kernel_launch(stream, kernel, triton_args);

  auto result = disableProfiler();
  result->save("./some_local_file2.json");

[anmyachev]

To avoid getting an incomprehensible error.

[kballeda]

I believe passing argc/argv to another function this way may not be a good idea.
we could use a parsing library to avoid this manual processing of various options, just as an example something like this (my past work).
https://github.com/Xilinx/Vitis_Libraries/blob/3c8a3d59fe76157ffa0b31fb267961bea2e148f5/data_compression/L3/demos/libzso/src/host.cpp#L271

we could use std::find instead of loop to search for the option

[kballeda]

as get_kernel_time is passed across launch(), sycl_launch_kernel() would it be possible to set it as part of constructor of struct ?

[kballeda]

message typo (NITS)