clfft on CPU and reusing host memory for data

[tdd11235813]

OpenCL allows to use a pointer to previous allocated buffer when cl_mem object is created with clCreateBuffer.
As long as padding is not required, the clfft benchmark on CPUs possibly could be speed up as upload/download is not needed, so have to check this when time permits.
(The benchmark suite itself uses the same buffer for input and output.)