Queue operations on the current device, not the one owning the array.

[maleadt]

When using host or shared buffer types, there isn't a clear owner, so let's use the CUDA semantics of executing on the currently-active device. This makes it possible to properly synchronize before copying host memory, which is needed for correctness.

[codecov]

Codecov Report

Attention: Patch coverage is 90.54054% with 7 lines in your changes missing coverage. Please review.

Project coverage is 81.22%. Comparing base (260a4dd) to head (ec6b79d).

Files	Patch %	Lines
lib/mkl/wrappers_blas.jl	90.47%	4 Missing ⚠️
lib/mkl/wrappers_lapack.jl	78.57%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #460      +/-   ##
==========================================
- Coverage   81.25%   81.22%   -0.03%     
==========================================
  Files          45       45              
  Lines        2571     2573       +2     
==========================================
+ Hits         2089     2090       +1     
- Misses        482      483       +1     

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