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

* fix wrong argument order in solve_impl call

solve_impl(n, x.ptr(), b.ptr());  // WRONG!
  The signature is solve_impl(int n, Real* b, Real* x), but x and b are swapped. This means the solver writes the solution to the wrong buffer. Should be:
  solve_impl(n, b.ptr(), x.ptr());

* fix Wrong buffer used for CPU float path

checksolver(cusolverSpScsrcholBufferInfoHost( handle, rows, nnz, descr, device_values, hRow, hCol, ...));
  This uses device_values instead of hValues (unlike the double version at line 124). This would access invalid memory when running on CPU with float precision.

Author: fredroy

src/SofaCUDALinearSolver/CUDACholeksySparseSolver.inl

@@ -126,7 +126,7 @@ void CUDASparseCholeskySolver<TMatrix, TVector>::setWorkspace()
         }
         else
         {
-            checksolver(cusolverSpScsrcholBufferInfoHost( handle, rows, nnz, descr, device_values, hRow, hCol,
+            checksolver(cusolverSpScsrcholBufferInfoHost( handle, rows, nnz, descr, hValues, hRow, hCol,
                 host_info, &size_internal, &size_work ));
         }
     }
@@ -434,7 +434,7 @@ void CUDASparseCholeskySolver<TMatrix,TVector>::solve(Matrix& M, Vector& x, Vect
         // LL^t y = Pb
         sofa::helper::ScopedAdvancedTimer solveTimer("Solve");
 
-        solve_impl(n, x.ptr(), b.ptr());
+        solve_impl(n, b.ptr(), x.ptr());
         checkCudaErrors(cudaStreamSynchronize(stream));
     }