Host-side stack unwinding

[brain-dump of potentially interesting approach for exception traces]

The exception backtrace POC in https://github.com/JuliaGPU/CUDAnative.jl/pull/303 embeds debug information in global variables and reports it at run-time since we don't have the ability to unwind the stack on-device. At the same time, unwinding the stack should be possible from the host, using the [CUDA debugger API](https://docs.nvidia.com/cuda/debugger-api/r_main.html#r_main) (this requires https://github.com/JuliaGPU/CUDAnative.jl/issues/31, of course). The only remaining problem there is how to transfer control from the device to the host once an exception occurs, since running under the debugger all time is probably expensive. the `CUDA_DEVICE_WAITS_ON_EXCEPTION` environment variable might be a possibility: https://docs.nvidia.com/cuda/cuda-gdb/index.html#attaching