Improve CUDA device detection

[makortel]

In #100 when SwitchProducerCUDA checks whether there are any CUDA devices on the system, it requires that all devices must have compute capability above a certain level. If any device is below, the switch decides that there are no CUDA devices.

There are many ways how this logic could be improved

• Disable the use of only those devices under the threshold.
  • Would require communication of the devices to CUDAService, or duplication of the logic
• Figure out the supported architectures from SCRAM CUDA flags

[fwyzard]

An other option:

• Try to run a simple kernel, and if that fails (because we did not compile for that gpu architecture) drop the device from the list of available device
  • this may try to initialise a device that we would otherwise avoid, and/or before it has been properly configured ?
  • to address that, is it enough to call cudaDeviceReset() before the explicit configuration and initialisation in the CUDAService ?

[fwyzard]

See #286 .