Add `cuda::std::abort` for uniform, heterogeneous program termination

As an author of CUDA C++ code, I often need to handle erroneous code paths. Oftentimes I'm not concerned with "proper" error handling and I just want crash the program without regard for anything else.

A variety of ways exist to do this in host code, but my options are more limited in device code.

The usual best practice in device code is [`__trap()` ](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#miscellaneous-instructions-trap). Note: this will corrupt the CUDA context and require restarting the process.

It would be nice to have an implementation of `cuda::std::abort` that handles calling `__trap()` when in device code and aborting in the usual way in host code.