[Performance] To use a proper default `num_warps` and `threads_per_warp` for Intel GPU backend

[chengjunlu]

The default number of the num_warps=4 and threads_per_warp=32 are come form the NV GPU. It is not optimal for Intel GPU.

We need use a proper default value based on GPU arch, double GRF and related setting for Intel GPU.

[vlad-penkin]

Measurements are required to prove that new defaults will work better