* Use common TextureInfo struct for all devices, except CUDA fermi. * Move image sampling code to kernels/*/kernel_*_image.h files. * Use arrays for data textures on Fermi too, so device_vector<Struct> works.