fix: update to reflect the fact that coin's learn does not return a prediction

[jackgerrits]

The prediction it currently returns essentially represents what the prediction would have been post learn and therefore does not match the contract of learn_returns_prediction.

[ataymano]

Essence of the issue:
data.txt:

1 | x:1
1 | x:1
1 | x:1
-100 | x:10

vw -d data.txt -p pred.txt --quiet --coin --noconstant

pred.txt:

0
1.000000
1
0.181818

The fact that last prediction got decreased indicates that it is not pre-learn predict but the one affected by learning on -100.
Although looks like this change is not enough and final prediction file still get overwritten prediction from learn stack

[ataymano]

Just checked manually - example from above got passed.