compute_device_list is using static vector of device information which
had pointers (identifier and name) to values from device information
structures. That structures are also stored in static vector and being
refreshed every 5 seconds.
The issue is, as soon as device information is being updated, pointers
in vector from compute_device_list became incorrect.
Seems it was the reason of issues with sudden switching from CUDA to
OpenCL on my desktop and from CUDA to CPU on my laptop, It was also
seems to be making persistent images behaves instable.
Made it so device identifier and name are copied from device info
to structures used by RNA (CCLDeviceInfo).
Alternative could be avoid cacheing CCLDeviceInfo and always use actual
list of device information by RNA. It shouldn't be so much slow.
The rendering device is now set in User Preferences > System, where you can
choose between OpenCL/CUDA and devices. Per scene you can then still choose
to use CPU or GPU rendering.
Load balancing still needs to be improved, now it just splits the entire
render in two, that will be done in a separate commit.
