AnalysisHandler cleanups

This change set

• adds an effNumEvents() in the AH (cf. discussion in #288 (closed))
• syncs the TMP-AO treatment with 3-1-x
• fixes the logic for sumW() and sumW2(): It should return the value for the currently active weight and only fall back to the nominal value of no active weight is set
• adds support for multithreading in rivet-merge via a multiprocessing pool on the Python side with pure in-memory merging of the resulting sub-steps
• adds a YODA::BinnedEstimate<int> to the AnalysisHandler following the discussion in #266 (closed). This AO has the beam species as its discrete edges and the beam z-momenta as the bin values (the convention is that the first bin is the beam with negative z-momentum, the second the one with positive beam momentum). It's not multi-weighted, but it is written out to the output stream and integrated into the (de-)serialization methods.
• introduces a --assume-reentrant flag to the rivet-merge script that reproduce the 3.1.x behaviour. Otherwise all non-reentrant-safe routines are skipped to ensure the merged result is well defined. An optional feature requested by @stefan-kiebacher in this discussion.