lack of memory pool in ctypes cuda driver

Comparing to cupy, a loop over frames is worse on the driver because it is doing a realloc each time. cupy used a pool. There is a different cuda alloc/free option that is better for realloc too.

Fixed in #46 on recent cards. For older cards check if there is an out= arg in the conveyor pipeline (and that we have that on CPU, the old API was set up to be re-using memory by default)