PyTorch profiler on all platforms

[hatanp]

Validating and improving how profiling is done. Internal Ipex profiling example works in isolation, but with this Megatron-DeepSpeed we are missing the XPU outputs.

• [ ] Timeline output with XPU, including kernels and communication
• [ ] Compare support and output of the PyTorch legacy profiler
• [ ] Check whether IPEX_ZE_TRACING=1 is needed and add that to relevant scripts
• [ ] Generate smaller profiler output files by leveraging torch.profiler.schedule and profiler.step(). For example wait 2 steps, warmup for 2 steps, and then profile 2 steps and write the output.

[hatanp]

Intel guide can be found at https://intel.github.io/intel-extension-for-pytorch/xpu/latest/tutorials/features/profiler_kineto.html Notable are the known issues and env variables.