This commit changes the memory layout to be lane-interleaved, so that
we can map only the necessary parts of GPU memory with OpenCL.
Previously we mapped the whole memory buffer, which was slow as hell...