Binning inefficiency

[jvansanten]

Dashi histograms drop down to numpy.histogramdd() on every call to fill(). For histograms with a large number of bins, this creates two inefficiencies:

1. Bin indices are found by binary search, when they're usually uniform under some transformation (i.e. linear or log)
2. A copy of the bin content array is created afresh on every call. Repeated calls to fill() would be much more efficient if bin content were added in place.

Things to do:

1. Create a binning object that implements the index calculation
2. Make these manipulable in the same way as bin edges (rebinning, slicing, etc)
3. Serialize and deserialize to HDF5 files
4. Reimplement fill() in terms of index calculation