Introduce variant based on cuDSS

[fredroy]

Based on

• Fix solve_impl() mixups args and cusolverSpScsrcholBufferInfoHost() call with wrong values #10

Add a new CUDASparseCholeskySolver variant using cuDSS: https://developer.nvidia.com/cudss

Needs cuda toolkit 12 and cudss to be installed (not included in the toolkit itself)

Benches: (FEMBAR_CUDASparseCholeskySolver* scenes)

legacy:500 iterations done in 17.3432 s ( 28.8298 FPS)
cudss: 500 iterations done in 4.83628 s ( 103.385 FPS)

same with a bigger grid 76x16x16

legacy:100 iterations done in 55.1384 s ( 1.81362 FPS)
cudss :100 iterations done in 15.4386 s ( 6.47726 FPS)

+ clean up the cmake file.
+
While doing this PR, I noticed the typo on the filenames of the legacy solver "CUDACholeksySparseSolver" 🥴

[alxbilger]

I managed to get it working.

For the record:

• I defined cudss_DIR to C:\Program Files\NVIDIA cuDSS\v0.7\lib\13\cmake\cudss in CMake
• The API changed from cudss 0.7 to 0.8. It fails to compile with 0.8. Let's merge it for now and upgrade later.
• The plugin could not be loaded because the cudss DLL was not found. I manually copied it next to the SofaCudaLinearSolver DLL. Maybe something to do with SofaCUDALinearSolverConfig.cmake.in?