Implement Barzilai Borwein step size rule

[jakobsj]

The Barzilai-Borwein step size very often gives much faster convergence and is cheaper to evaluate than simple backtracking. It should be implemented in gradient descent.

[MargaretDuff]

https://en.wikipedia.org/wiki/Barzilai-Borwein_method

Main reference is the first one: Barzilai, Jonathan; Borwein, Jonathan M. (1988). "Two-Point Step Size Gradient Methods". IMA Journal of Numerical Analysis. 8: 141–148

https://doi.org/10.1093%2Fimanum%2F8.1.141