* Reduce unused macros, where two strategies are used:
(1) Extract batch_headers and let them be included only once in each batch.
(2) Static reduction in primitive_db_gen.py, which scans each macro's users and exclude the macro if there is no user.
* Removed dependency from runtime to kernel_selector
* Resolve too large string error
* Fix duplicated definition (GET_FILTER_XXXX is defined in both fetch_weight.cl and by runtime.
Removed from runtime because the definition is incorrect
* Resolve GRN & deconv & gpu_select issues
* Fix cldnn unittest issues
* Minor fix
* Applied review comments
* Fix rebase error