don't fail if profiler already initialized

[michael-kenzel]

Instrumentation calls cuptiProfilerInitialize to enable higher precision GPU timestamps. However, we currently fail if the profiler was already initialized, which prevents running with nvprof. We should fix this by not failing if cuptiProfilerInitialize fails due to the profiler already having been initialized.

[richardmembarth]

This fails also on my machine without running nvprof.
Just running any CUDA binary fails with error code 15: CUPTI_ERROR_NOT_INITIALIZED

[richardmembarth]

Maybe, we should check for this error code and continue without CUPTI.

CUPTI calls will fail with error code CUPTI_ERROR_NOT_INITIALIZED if the CUDA driver version is not compatible with the CUPTI version.

https://docs.nvidia.com/cupti/r_main.html