Fix error in BieglerKKTSystem's solve for nontrivial scaling

[frapac]

cc @amontoison : that explains why we observed a large errors in the KKT benchmark.

[codecov]

Codecov Report

Attention: 8 lines in your changes are missing coverage. Please review.

Comparison is base (e81bbbd) 76.76% compared to head (cde3740) 69.11%.

Files	Patch %	Lines
ext/api.jl	0.00%	8 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #71      +/-   ##
==========================================
- Coverage   76.76%   69.11%   -7.65%     
==========================================
  Files          22       28       +6     
  Lines        2152     2409     +257     
==========================================
+ Hits         1652     1665      +13     
- Misses        500      744     +244     

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

[michel2323]

Looks good to me. So ] test now uses CUDA 12.1 for now.

[amontoison]

Can we use CUDA 12.3 by default?

[michel2323]

I should mention that ] test fails in CUSOLVERRF with

  The current version of CUDA is not supported.

[amontoison]

Yes, we only support <= v11.8 or >= v12.3 for the version of CUDA in CUSOLVERREF.jl.

[frapac]

Can we use CUDA 12.3 by default?

Right now, I prefer to keep CUDA runtime 11.8 as the default, as this is the one we are using for our benchmarks. We should be able to replace cusolverRF by CUDSS soon, but I don't know if CUDA 12.3 is available on polaris, which could be a problem for us.

An alternative solution is to test both for CUDA 11.8 and 12.3 in Argos. But that could be addressed in a separate PR.