Operations inside for loops are unnecessary. They are repeated anyway just below for loops

Core_ArithmMask.uninitialized

Add new ocl kernel for image pyramids upsampling,
It is 35% faster than current OCL kernel in perf test.
Signed-off-by: Li Peng <peng.li@intel.com>

Usage:
    writer.set(cv::CAP_PROP_IMAGES_BASE + cv::IMWRITE_JPEG_QUALITY, 80);