Add adaptive eigendecomposition computation frequency for eigh

[runame]

Just like (approximate) eigendecompositions can be skipped in the warm-started QR algorithm, we can skip eigendecompositions with torch.linalg.eigh using the same criterion.

This PR implements this for EigendecomposedShampooPreconditionerList and EigenvalueCorrectedShampooPreconditionerList.

[tsunghsienlee]

It seems the best place to implement this is inside matrix_eigendecomposition() with only adding this field into EigendecompositionConfig from the interface/engineering perspective (not sure this is the best from the algorithm/semantic level, need to discuss this with you). If this is possible, following are the benefits:

1. Limit this config so we don't need to worry about this for ShampooPreconditionerConfig.
2. No need to modify shampoo_preconditioner_list.py.
3. Keep _eigenvalues_estimate_criterion_below_or_equal_tolerance() as it is now, and people who use matrix_eigendecomposition() could still use this always.