No service fee on penalties

[fhenneke]

With this PR, if a solver receives a negative reward for an accounting week, that amount is not multiplied by reward_scaling.

Before the change, negative rewards were multiplied by reward_scaling independent of the sign of the reward. For positive rewards, this does what it should, as a fraction of the reward is withheld from the solver and becomes part of a bounty budget. For negative rewards, the calculation implied that part of the penalty is withheld and it reduces the bounty budget.

With the change, negative batch rewards are not multiplied by reward_scaling. Quote rewards are left as is.

The proposed implementation changes some code in 3 places. Ideally there would only be one change required. Note, that quote rewards are always non-negative and no change in the formula is required.

Alternatively, the total cow reward could be used to decide on the scaling, instead of handling batch and quote rewards separately.

Tests have not been adapted yet.