[libomptarget] Implement atomic inc and fence functions for AMDGCN using clang builtins

This function uses __builtin_amdgcn_atomic_inc32():
  uint32_t atomicInc(uint32_t *address, uint32_t max);

These functions use __builtin_amdgcn_fence():
__kmpc_impl_threadfence()
__kmpc_impl_threadfence_block()
__kmpc_impl_threadfence_system()

They will take place of current mechanism of directly calling IR functions.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D83132