Refactor MSR Banzhaf valuation

[janosg]

Description

This PR implements Maximum Sample Re-use (MSR) Banzhaf valuation in the new architecture. The implementation deviates strongly from the previous implementation and fixes a bug in the variance estimation.

The new implementation uses two ValuationResult instances to keep track of the positive and negative running means. After each update, those are combined into the final result object. The update counter of the combined result is set to the minimum of the two update counters. The variance of the combined result is set to the sum of variances (assuming independence).

Open questions

• How should update counters for MSR be defined? I chose the minimum of the positive and negative update counters because this makes the counters comparable to the counters for normal semivalues.
• To which value should we set the value estimates if either the positive or the negative mean had zero updates? The paper suggests zero but I think nan is safer. Note that if the update counter is defined as suggested above, this situation can be completely avoided by using MinUpdates as stopping criterion.
• To which value should we set the variance estimates if either the positive or the negative mean had zero updates? I chose inf.

Checklist

• [x] Wrote Unit tests (if necessary)
• [x] Updated Changelog

[janosg]

I clarified the documentation of ValuationResult.variance, but I still think it is a bit misleading that ValuationResult.variances is not just the square of ValuationResult.stderr.

I think it would be clearer if we only expose the square root of the variances as ValuationResult.stdev; Then it's clear that the difference betweenn stdev and stderr must be a conceptual one. Also, most of the time standard deviations are more interpretable than variances.