CUDA device selection by name

[Antoine Martin]

Follow up for #2415: we should be able to select the list of enabled / disabled devices using their name and / or PCI bus location (ie: 0000:1A:00.0) without doing a full CUDA context initialization.

load_device can get to the device name before calling make_context (which is slow).

[Antoine Martin]

We can't skip the make_context completely in select_device because driver.mem_get_info() requires a valid context.

Maybe we can have different policies for load balancing, and one may not require memory info to do its job.

[Antoine Martin]

Updates:

• r24331: can use device name or PCI id to enable or disable devices
• r24332 + r24333: update examples (maybe we should ignore the PCI prefix? or allow a pattern match?)

[Antoine Martin]

r24340 adds a load-balancing attribute and a round-robin option, the memory policy is still the default. (chooses the GPU with the most free memory)

Still needs testing.

[Antoine Martin]

Seems to have caused a bug: #2489.

See also initial load balancing code in #520.