clang-tidy review says "All clean, LGTM! 👍"

cc: @guitargeek. I have tested this with the atlas likelihood benchmark code, it now works after removing the auxiliary arrays. The runtime results have also improved because of this.

The failed tests are due to the CUDA download server not being available.

The issue was that while calling pullbacks, we don't maintain non-differentiable arguments. @PetroZarytskyi had fixed this for most of the cases here: #802, but it still had some more complex cases which required activity analysis, as he mentioned in the PR too.
My PR just tries to improve upon that by handling the case for const pointers, which can be handled without activity analysis too.

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 94.08%. Comparing base (9579ce4) to head (3516cbc).

@@           Coverage Diff           @@
##           master     #919   +/-   ##
=======================================
  Coverage   94.07%   94.08%           
=======================================
  Files          53       53           
  Lines        7747     7757   +10     
=======================================
+ Hits         7288     7298   +10     
  Misses        459      459           

Looks good to me. Sure, this will probably become obsolete once we have activity analysis but this is a great simplification for now.