Quick fix to multithreading issue by refining specialization of contravariant vectors

[tristanmontoya]

This is an attempt to fix the problem @benegee noted in #61, in which the cases which use "standard" Trixi.jl solvers (i.e. not those for PDEs on surfaces) return an error when multiple threads are used, due to TrixiAtmo.jl's specialized get_contravariant_vector method for PtrArray types being called inadvertently. I'm not super familiar with the workings of PtrArray, but looking at the StrideArraysCore.jl code, I was able to figure out what type parameters would be associated with a 2D mesh in 3D space, and therefore hard-coded the dispatch for those dimensions. With this change, the issues in #61 are no longer encountered.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.45%. Comparing base (e7c1de7) to head (5af6773).
Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #62      +/-   ##
==========================================
- Coverage   89.46%   89.45%   -0.01%     
==========================================
  Files          20       20              
  Lines        1803     1802       -1     
==========================================
- Hits         1613     1612       -1     
  Misses        190      190              

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

[benegee]

Great fix. Thanks a lot @tristanmontoya !
I am not super familiar with this either, but obviously your specialization works and it contains NDIMS as well as NDIMS_AMBIENT, so it cannot interfere with Trixi's function anymore. The only concern would be that the type is too specific now and does not cover all your cases. If this not an issue, we could go for it.

[tristanmontoya]

Great fix. Thanks a lot @tristanmontoya ! I am not super familiar with this either, but obviously your specialization works and it contains NDIMS as well as NDIMS_AMBIENT, so it cannot interfere with Trixi's function anymore. The only concern would be that the type is too specific now and does not cover all your cases. If this not an issue, we could go for it.

Thanks! It covers all the cases we use now, and doesn't have a significant impact on performance. We can look for another solution if needed later on.

[ranocha]

It would be great to add some CI tests running an example with multiple threads to ensure that the bug does not occur anymore