Delta learning doesn't work in Mean-Approx mode

[eric-tramel]

When using a full, explicit, matrix, ample gives correct performance in the following cases:

• Known noise and known prior parameters.
• Known noise and unknown prior parameters.
• Unknown noise and known prior parameters.
• Unknown noise and unknown prior parameters.

However, when a mean-approximation is used for the F2(...) and F2T(...) operations, delta learning breaks down, working in the following cases:

• Known noise and known prior parameters.
• Known noise and unknown prior parameters.
• Unknown noise and known prior parameters.
• Unknown noise and unknown prior parameters.

This may relate to the manner in which Delta is being updated.