[rossjjennings]

This is a revised implementation of Archive.tscrunch() that uses np.add.reduceat() (see Numpy docs, also StackOverflow) to avoid the slow nested for loops pointed out in #16.

It also gets rid of a bug in the original tscrunch() implementation where, if the factor didn't divide evenly into the number of subintegrations, the "leftover piece" would just get filled with zeros, never with actual data. For example, if there are 12 subintegrations in the original archive and factor=5, the final scrunched archive should have 3 subintegrations, where the new subint 0 is the average of old subints 0, 1, 2, 3, and 4, new subint 1 is the average of old subints 5, 6, 7, 8, and 9, and new subint 2 is the average of old subints 10 and 11. In the old version, new subint 2 would be all zeros, but here it actually gets filled in.

I have tested against the old implementation on a .zap file, and the fractional differences are at the 10^-16 level, consistent with rounding error from doing the additions in a different order -- except, of course, for the "leftover piece". The new implementation took about 172 ms in my testing, as opposed to about 1 minute for the original.