New algos are faster / more accurate

[tdunning]

The MergingDigest in the java reference implementation is considerably fast than the AVL tree implementation and has a number of accuracy tweaks that really help.

Some of the major wins include:

• no dynamic allocation at all.

• stratified merging gives 5-10x better accuracy

• better interpolation makes accuracy better again

• multiple scale functions improve accuracy and speed, all uniform accuracy mode

• firm bounds on digest size for all scale functions. Bounds are also nearly consistent across different scale functions.

I can help with the upgrades if you like.