FR: add helpful error messages for inadequate-cuda-capability detection

[steven-johnson]

Currently, if you compile for (say) host-cuda-cuda_capability_50, but run on a system that only supports lower versions of Cuda, you'll fail at runtime with a PTX_INVALID failure, which isn't very helpful. It would be nice if the runtime assertion code we generate sniffed the version-found vs the version-required and emitted a more useful error message.