[GPU] POC for new OCL kernels approach

[vladimir-paramuzov]

Details:

• Currently we have separate kernel selector component with independent API and kernels list which leads to some data structures conversion overheads and it doesn't allow to have unified kernel selection logic. So the goals of this change are:
  • Make OCL jitters work with common primitive/layout descriptors
  • Have better multi-kernel impls design
  • Allow registration of OCL impls in the common impls list
  • Use ImplementationManager to check applicability of individual OCL impl
• In the scope of this PR few impls are updated: CM LSTM kernel (and removed old infra) and CTC Loss kernel from OCL
• Enable clang-format for the updated backends

[yeonbok]

random spot:
It is little bit not intuitive for me that the kernels_cache & header cl kernels are under impls, while the cl kernels are under kernels_selector directory, and impls and kernels_cache are in the same hierarchy..
(also there are header cl kernels under kernels_selector directory) ..
What was the main idea of the new directory construction? Could you add some comment regarding the criteria?

[vladimir-paramuzov]

The main idea is that backend specific code should be located in impls/${backend_name} folder. headers for kernel selector are in kernel selector still (those are duplicated for now to avoid any dependencies between components). Kernel selector shall be removed in the future, so the whole structure of ocl backend will look like ocl_v2 version

[jkasprza]

GEMM kernel should be executed before loop kernel here.