Use unified helper functions for creating and destroying OpenCL buffers/mappings - and/or for re-introducing CUDA

Many formats have variants of this:

#define CLCREATEBUFFER(_flags, _size)         \
    clCreateBuffer(context[gpu_id], _flags, _size, NULL, &cl_error); \
    HANDLE_CLERROR(cl_error, "Error allocating GPU memory");

#define CLKERNELARG(kernel, id, arg)      \
    HANDLE_CLERROR(clSetKernelArg(kernel, id, sizeof(arg), &arg), \
                   "Error setting kernel argument");
(...)

We should have such macros in opencl_common.h and use them everywhere. They can have memory-leak detection like my debug macros for #4354 (could in fact be in place all the time, not just for debugging).

Start by reading opencl_common.h and comparing the format-specific buffer and mapping macros described in the issue. Review the debug macros from #4354 and determine the intended common interface and leak-detection behavior. Done means the relevant OpenCL paths use unified helpers, with any CUDA reintroduction scope explicitly settled.