automated tests should be more robust now

In previous releases, the tests run by travis (tests/test_jacobi.cpp) only work if the arguments supplied are chosen to avoid large differences in eigenvalue magnitudes. Due to numeric underflow, it is normal for significant errors to appear in low-magnitude eigenvalues (and corresponding eigenvectors) whenever the difference between the magnitudes of the smallest and largest eigenvalues exceed 1.0e10. This is normal even if the algorithm is working correctly. To be fair, huge differences in eigenvalue magnitudes like this are not usually encountered. However such matrices would cause the automated tests to fail. Now, the automated tests are smart enough to avoid concluding that the test has failed in these cases.

A new argument (eps) was also introduced allowing the tolerance to be set by the caller (instead of a constant). (It is 1.0e-06 by default.)