Interface with JHUGen/MELA

[lk11235]

I should add to the setup script some kind of automatic compilation of JHUGen/MELA with python-version compatible interface. Note that the MELA interface was built in python2 but should work ootb with python3 after the latest fixes.

One solution to using MELA in the Tagger could be to separate the MELA and the discriminant parts from the main Tagger. Speed-up of new probability calculations will likely be accomplished with vectorized operations over immutable arrays of MELA objects, so after using MELA, the Tagger can pass the final numbers to the discriminant calculator (which should also handle event tagging). Then the Tagger takes care of writing, merging, and cleaning the trees.

[lk11235]

Leading suggestion (for simplicity and efficiency) is to run MELA calculations and discriminant calculations in different subprocesses. Each can be vectorized and batches of calculations can be parallelized. But XLA may be a cleaner approach (parallelize operations rather than unique calculations).