This reverts commit 7d91b8ef.

Fix problems on NVidia devices.

Update according to vbystricky's comments

For current OpenCV-CL architecture, if the data buffer
allocated in UMat are cpu buffer(not ocl buffer) under
cpu mode, and then pass this UMat to an OpenCL kernel
as an argument, the OpenCL path will fail and fallback
to cpu mode. Take HOGDescriptor::oclSvmDetector as an example:
    ocl::setUseOpenCL(false);
    //data allocated in hog.oclSvmDetector will be cpu buffer
    hog.setSVMDetector(HOGDescriptor::getDaimlerPeopleDetector());
    ocl::setUseOpenCL(true);
    //We enabled OpenCL, but hog.oclSvmDetector are cpu buffer,
    //so it will fail in the function ocl_classify_hists
    //when reach to this line
    //idx = k.set(idx, ocl::KernelArg::PtrReadOnly(detector));
    hog.detectMultiScale(img, found, hit_threshold, win_stride,
            Size(0, 0), scale, gr_threshold);

Similar problems heppen on img_aux and img. So we should re-define
or re-set these UMat when do mode switch (CPU -> OpenCL) in order
to make their data be allocated by ocl and then OpenCL path will
succeed.
Signed-off-by: Chuanbo Weng <chuanbo.weng@intel.com>