Make tridiagonal QR algorithm more precise

[andreasnoack]

For some reason, my implementation requires a much smaller tolerance than LAPACK to get the same precision in the eigenvectors. I'm not sure why. Things to check

1. Will a QR/QL strategy be more precise? Right now I'm only doing QR.
2. Is the shift calculation avoiding canceling operations?

[andreasnoack]

I think this has improved but original poster should have included a reproducer