quad_form_diag derivatives

[spinkney]

Provide a reverse-mode specialization for quad_form_diag.

[bob-carpenter]

I rewrote the issue as an issue and tagged it as a feature and good first issue, but it's still missing all the boilerplate.

Is it faster or negligibile to just autodiff through the diag version?

A reverse specialization would be faster and more memory efficient. It probably hasn't been optimized because it's mainly used for rescaling a covariance matrix, which we'd rather do in the Cholesky space for efficiency.