Update Spectra to v1.0.1 to fix C++20 compatibility issues on GCC/Clang

[valgur]

GTSAM currently unfortunately fails to build with C++20 or newer on GCC and Clang with the following errors in ShonanAveraging.cpp:

  [ 70%] Building CXX object gtsam/CMakeFiles/gtsam.dir/sfm/ShonanAveraging.cpp.o
  In file included from /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsBase.h:20,
                   from /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsSolver.h:12,
                   from /home/martin/libs/gtsam/gtsam/sfm/ShonanAveraging.cpp:19:
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:108:50: error: expected ‘)’ before ‘*’ token
    108 |     ArnoldiOp<Scalar, OpType, IdentityBOp>(OpType* op, IdentityBOp* /*Bop*/) :
        |                                           ~      ^
        |                                                  )
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsBase.h: In instantiation of ‘Spectra::SymEigsBase<Scalar, SelectionRule, OpType, BOpType>::SymEigsBase(OpType*, BOpType*, Index, Index) [with Scalar = double; int SelectionRule = 0; OpType = gtsam::MatrixProdFunctor; BOpType = Spectra::IdentityBOp; Index = long int]’:
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsSolver.h:165:83:   required from ‘Spectra::SymEigsSolver<Scalar, SelectionRule, OpType>::SymEigsSolver(OpType*, Index, Index) [with Scalar = double; int SelectionRule = 0; OpType = gtsam::MatrixProdFunctor; Index = long int]’
  /home/martin/libs/gtsam/gtsam/sfm/ShonanAveraging.cpp:641:79:   required from here
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/SymEigsBase.h:252:15: error: no matching function for call to ‘Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>::ArnoldiOp(gtsam::MatrixProdFunctor*&, Spectra::IdentityBOp*&)’
    252 |         m_fac(ArnoldiOpType(op, Bop), m_ncv),
        |               ^~~~~~~~~~~~~~~~~~~~~~
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note: candidate: ‘constexpr Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>::ArnoldiOp(const Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>&)’
     99 | class ArnoldiOp<Scalar, OpType, IdentityBOp>
        |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note:   candidate expects 1 argument, 2 provided
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note: candidate: ‘constexpr Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>::ArnoldiOp(Spectra::ArnoldiOp<double, gtsam::MatrixProdFunctor, Spectra::IdentityBOp>&&)’
  /home/martin/libs/gtsam/gtsam/3rdparty/Spectra/MatOp/internal/ArnoldiOp.h:99:7: note:   candidate expects 1 argument, 2 provided
  gmake[2]: *** [gtsam/CMakeFiles/gtsam.dir/build.make:2050: gtsam/CMakeFiles/gtsam.dir/sfm/ShonanAveraging.cpp.o] Error 1
  gmake[1]: *** [CMakeFiles/Makefile2:1397: gtsam/CMakeFiles/gtsam.dir/all] Error 2
  gmake: *** [Makefile:166: all] Error 2

This PR updates the vendored Spectra from v0.9.0 to the latest v1.0.1 as a simple fix for these C++20 compatibility issues. I also moved the Spectra headers under a Spectra/... prefix to match the installed header paths since v0.7.0.

Change log: https://github.com/yixuan/spectra/blob/master/CHANGELOG.md

Since this is a major version bump, there are some breaking changes in the Spectra API, which are listed in the migration guide: https://spectralib.org/upgrade

Corresponding changes to ShonanAveraging.cpp:

diff --git a/gtsam/sfm/ShonanAveraging.cpp b/gtsam/sfm/ShonanAveraging.cpp
index 7c8b07f37c..151c816b21 100644
--- a/gtsam/sfm/ShonanAveraging.cpp
+++ b/gtsam/sfm/ShonanAveraging.cpp
@@ -16,7 +16,7 @@
  * @brief  Shonan Averaging algorithm
  */
 
-#include <SymEigsSolver.h>
+#include <Spectra/SymEigsSolver.h>
 #include <cmath>
 #include <gtsam/linear/PCGSolver.h>
 #include <gtsam/linear/SubgraphPreconditioner.h>
@@ -574,6 +574,8 @@ static bool PowerMinimumEigenValue(
  * nontrivial function, perform_op(x,y), that computes and returns the product
  * y = (A + sigma*I) x */
 struct MatrixProdFunctor {
+  using Scalar = double;
+
   // Const reference to an externally-held matrix whose minimum-eigenvalue we
   // want to compute
   const Sparse &A_;
@@ -634,13 +636,13 @@ static bool SparseMinimumEigenValue(
     Eigen::Index numLanczosVectors = 20) {
   // a. Estimate the largest-magnitude eigenvalue of this matrix using Lanczos
   MatrixProdFunctor lmOperator(A);
-  Spectra::SymEigsSolver<double, Spectra::SELECT_EIGENVALUE::LARGEST_MAGN,
-                         MatrixProdFunctor>
-      lmEigenValueSolver(&lmOperator, 1, std::min(numLanczosVectors, A.rows()));
+  Spectra::SymEigsSolver lmEigenValueSolver(
+      lmOperator, 1, std::min(numLanczosVectors, A.rows()));
   lmEigenValueSolver.init();
 
-  const int lmConverged = lmEigenValueSolver.compute(
-      maxIterations, 1e-4, Spectra::SELECT_EIGENVALUE::LARGEST_MAGN);
+  const int lmConverged =
+      lmEigenValueSolver.compute(Spectra::SortRule::LargestMagn, maxIterations,
+                                 1e-4, Spectra::SortRule::LargestMagn);
 
   // Check convergence and bail out if necessary
   if (lmConverged != 1) return false;
@@ -668,10 +670,8 @@ static bool SparseMinimumEigenValue(
 
   MatrixProdFunctor minShiftedOperator(A, -2 * lmEigenValue);
 
-  Spectra::SymEigsSolver<double, Spectra::SELECT_EIGENVALUE::LARGEST_MAGN,
-                         MatrixProdFunctor>
-      minEigenValueSolver(&minShiftedOperator, 1,
-                          std::min(numLanczosVectors, A.rows()));
+  Spectra::SymEigsSolver minEigenValueSolver(
+      minShiftedOperator, 1, std::min(numLanczosVectors, A.rows()));
 
   // If S is a critical point of F, then S^T is also in the null space of S -
   // Lambda(S) (cf. Lemma 6 of the tech report), and therefore its rows are
@@ -699,8 +699,9 @@ static bool SparseMinimumEigenValue(
   // order to be able to estimate the smallest eigenvalue within an *absolute*
   // tolerance of 'minEigenvalueNonnegativityTolerance'
   const int minConverged = minEigenValueSolver.compute(
-      maxIterations, minEigenvalueNonnegativityTolerance / lmEigenValue,
-      Spectra::SELECT_EIGENVALUE::LARGEST_MAGN);
+      Spectra::SortRule::LargestMagn, maxIterations,
+      minEigenvalueNonnegativityTolerance / lmEigenValue,
+      Spectra::SortRule::LargestMagn);
 
   if (minConverged != 1) return false;

To be honest, I would personally be hesitant to merge such an unwieldy PR, so you might want to commit these changes yourself or at least diff them against the official Spectra sources.