Add `diag_fisher` functionality to allow the Hessian to be computed "on the fly"

[AlanTuQC]

Instead of storing the entire Hessian matrix, diag_fisher=True calculates Hessian rows only when they must be used (i.e. in the relevant step of coordinate descent). It accomplishes this by storing the current state of the 1-D array hessian_rows (which appears in the middle of the sandwich product). This necessitates a greater number of matrix operations (many of them redundant), but also saves memory.

To-Do:

• state.hessian does not need to be instantiated if diag_fisher=True; this is not yet accounted for in the current code.
• Make sure we've addressed all entry points where the argument diag_fisher can be set by the user.
• How do we implement memory views (and possibly nogil) in _cd_fast.pyx?