Can we run the verified integers on CUDA?

[mborland]

There's no throwing mechanism on device, but asserts will return a CUDA error code to the caller. This would require the user to check that. The other alternative would be to just __trap() in case the caller isn't checking their error codes.

[mborland]

With a synchronized error reporting mechanism we can print all the same diagnostic information, __trap(), and then the user will get useful information when they check the CUDA error reporting mechanism. I think we need to replace BOOST_THROW_EXCEPTION with something like BOOST_SAFE_NUMBERS_THROW_ERROR that is normally defined as BOOST_THROW_EXCEPTION, but can also dispatch to our CUDA specific error reporting mechanism.