Add new OpenCL kernels for bicubic interploation, it is 20% faster
than current warp image kernel with bicubic interploation.
Signed-off-by: Li Peng <peng.li@intel.com>