Add Shanno-Phua initial step size

[vladimir-ch]

• Implemented the other initial step size strategy described in Nocedal&Wright
• In some CG implementations the formula is attributed to Shanno&Phua, but for consistence with QuadraticStepSize the type's name FirstOrderStepSize rather refers to how the step size is estimated. ShannoPhuaStepSize is an alternative name anyway.
• This strategy seems to generate step size that work better with CG than those generated by QuadraticStepSize
• Improve QuadraticStepSize #16 should be reviewed and merged first

[coveralls]

Coverage increased (+0.12%) when pulling b0aca98 on first-order-step-size into 13d6e43 on master.

[btracey]

This needs a high-level comment about what it does.

[coveralls]

Coverage decreased (-2.24%) when pulling bd49a15 on first-order-step-size into a1b9141 on master.

[vladimir-ch]

PTAL @btracey

[coveralls]

Coverage decreased (-2.24%) when pulling d08d506 on first-order-step-size into a1b9141 on master.

[coveralls]

Coverage decreased (-2.26%) when pulling b73576a on first-order-step-size into 022ec44 on master.

[btracey]

This is backward.
... estimated at |g|_inf / InitialStepFactor

[vladimir-ch]

This is correct, we want to take a small step when the gradient is large (when it is small, the step size is bound by MaxStepSize).

[btracey]

LGTM with the comment changes.

[coveralls]

Coverage decreased (-2.69%) when pulling 3135309 on first-order-step-size into 022ec44 on master.

[coveralls]

Coverage decreased (-2.69%) when pulling 3135309 on first-order-step-size into 022ec44 on master.