Compatibility of 'DualNum' trait with nalgebra

[prehner]

By implementing nalgebra::ComplexField for dual numbers (analogous to the implementation for f64) the full linear algebra capabilities of nalgebra (in particular eigenvalues) can be used with arbitrary generalized (hyper-) dual numbers. Currently this is only possible with first order dual numbers.

In this context it is worth reconsidering whether the Float trait should also be part of the supertraits of DualNum.

If this is implemented, the linalg module can be discarded in favor of the well established functionalities in nalgebra.

[prehner]

The required traits for using the eigenvalue function in nalgebra are so unfathomably restrictive (e.g., including SimdValue) that it is unlikely that they will ever be implemented for structs other than f32 and f64. Therefore, it is unlikely that we can completely get rid of the eigenvalue algorithm in num-dual.

[prehner]

#59 actually does implement the necessary trait bounds that calculations like

    let a = Dual64::from(1.0).derivative();
    let b = Dual64::from(2.0);
    let c = Dual64::from(4.0);
    let m = Matrix2::from_row_slice(&[a, b, b, c]);
    let v = SymmetricEigen::new(m).eigenvalues;

are now possible. Therefore the linalg module and feature should be abandoned in favor of nalgebras tried and tested implementations.