Implement compute and I/O overlap in SMX

• Use a double buffering scheme to implement compute and I/O overlap of the pyFAI kernels.
• Make sure the transfer are DMA from host pinned memory.