Improve kkt creation on GPUs (#299)

* add wrapper to cuDSS solver

* initial implementation

* debug HybridCondensedKKTSystem

* undocumented cholesky solver added

* add support for GPU

* deactive iterative refinement for HybridKKTSystem

* update CUDSS interface

* clean testing scripts

* implement AMD ordering for CUDSS

* CUDSS: add inertia for LDL factorization

* remove permutation in cuDSS

* cudss: add safeguard for inertia

* Fix MOI interface

* metis

* metis

* option fixes

* option fix

* add wrapper to CHOLMOD

* added cholmod and ldl

* option issue fixed

* option fix

* cudss test changed

* addressed Francois' commeonts

* addressed Alexis' comments

* renamed options cudss

* added a few comments

* improved initialization

* buf fixes

* minor edits

* found a bug in scaling

* scaling improved

* sort error fix

* coo to csc bug fix

* addressed Francois' comments

* addressed Francois' comments

* name chanage and bug fix

* bug fix

---------

Co-authored-by: fpacaud <[email protected]>
Co-authored-by: Sungho Shin <[email protected]>

lib/MadNLPGPU/src/cusolverrf.jl

@@ -273,7 +273,6 @@ function MadNLP.factorize!(M::CuCholeskySolver)
         CUSOLVER.spcholesky_factorise(M.inner, M.fullp, eltype(M.fullp.nzVal) == Float32 ? 1e-6 : 1e-12)
         M.singularity = false
     catch e
-        println(e)
         M.singularity = true
     end