Add a concurrent kwarg to profiling macros.

[maleadt]

This allows switching between concurrent and serial profiling modes. The former has less of an impact on an application's performance characteristics, but requires kernel instrumentation. The latter doesn't, and generally has less overhead.

cc @thomasfaingnaert concurrent=false is a kwarg that should be accepted by all profiling macros. I haven't seen significant improvements, but as per NVIDIA this may depend on the type of kernel:

serial: For applications which use only a single CUDA stream and therefore cannot have concurrent kernel execution, this mode can be useful as it usually (not always) incurs less profiling overhead compared to the concurrent kernel mode.

concurrent: Due to the code instrumentation, concurrent kernel mode can add significant runtime overhead if used on kernels that execute a large number of blocks and that have short execution durations.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (8a0d39a) 72.93% compared to head (4e34ea6) 72.94%.
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2201      +/-   ##
==========================================
+ Coverage   72.93%   72.94%   +0.01%     
==========================================
  Files         159      159              
  Lines       14651    14643       -8     
==========================================
- Hits        10685    10682       -3     
+ Misses       3966     3961       -5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.