PCAModelBuilderWithSelfAdjointEigenSolverTest is failing [develop branch]

[muefra]

How to reproduce:

Here the difference is usually in the 7 figure range:

git checkout b30e7dc1d3c994351d4e6790d071a957f3fb59e0 && make -j $(nproc) && ctest -R PCAModelBuilderWithSelfAdjointEigenSolverTest -VV

Output:

5: PCAModelBuilderWithSelfAdjointEigenSolverTest:       comparing principal components...       [failed]
5: PCAModelBuilderWithSelfAdjointEigenSolverTest:       - pc error(5.24711e+09) exceeds threshold (150)!
5: PCAModelBuilderWithSelfAdjointEigenSolverTest:       comparing variances...          [failed]
5: PCAModelBuilderWithSelfAdjointEigenSolverTest:       - variance error(1.07573e+17) exceeds threshold (250)!

Here it's working (although not always since the threshold is 250, but it sometimes goes up to ~400)

git checkout b1a69864e1fcc51aa022a0404c7b5e06a8f63e67 && make -j $(nproc) && ctest -R PCAModelBuilderWithSelfAdjointEigenSolverTest -VV

Output:

5: PCAModelBuilderWithSelfAdjointEigenSolverTest:       comparing principal components...  (0.000409 sec)                       [passed]
5: PCAModelBuilderWithSelfAdjointEigenSolverTest:       comparing variances...  (0.000415 sec)                          [passed]

I have not yet had a look at what exactly causes the difference in PCA-variance vectors and samples to be so big between the Jacobi and the SelfAdjointEigenSolver models.