Store run average only for scoring meshes

[lnevay]

A placeholder issue that I will update as I make progress.

Currently, the scoring meshes are stored per-event. The run-level equivalent mesh is just a summation of all events and therefore not normalised and the statistical bin errors not correct. This also requires a very very large transient data set that is then processed by rebdsimHistoMerge to produce the run-average histograms that can be also combined properly using rebdsimCombine.

The request is to by default provide the correct run-average histograms in the output and only optionally store the per-event histograms. This would vastly reduce the transient data size that has become problematic for a lot of people.

[lnevay]

This requires linking (lib)bdsim to (lib)rebdsim to make use of the accumulators. This isn't a problem but should be noted for future better structuring / cmake of the project.

A new accumulator was written for 1-4D histograms that efficiently only accumulates bins that were filled. As usual, in the end of event action, we loop over hits map and fill the appropriate histograms. The ROOT TH1::Fill methods (overridden for TH1,2,3D) return a global bin index (integer). This is cached in a set (therefore no duplicates). A new mark the end of event function invokes the accumulation once all bins have been filled (potentially multiple times in the same bin). Only the bins that were filled are accumulated (avoids loop over whole histogram). A duplicate structure histogram is kept internally that counts in the same bin the number of events that were filled at all. So only +=1 even if in that event multiple hits were filled in one bin.

At the end of run, the accumulators are terminated. Here, we do one loop over every single bin (low overhead in comparison) and add on (n_events - n_events_in_bin) x 0 entries. An appropriate implementation of the online mean and variance calculation was implemented for this. This produces numerically identical results as the naive looping over all bins each event and accumulating the zeros, but is significantly faster.

[lnevay]

An overall consequence is that the run-level histograms are not the correct average histograms directly. e.g. for ElossHisto and ElossPEHisto. This is a major improvement. There are a few remaining tasks before the pull request is ready:

1. rebdsimHistoMerge must be altered to only copy the run-average histogram (only for the next overall data version upwards)
2. rebdsim must be altered in a similar fashion
3. options should be created to store the event-level histograms, i.e. we don't need event-level ElossHisto by default
4. algorithm needs to be documented in numerical methods as well as bin caching