Uproot

Fixed 102.

Simplified Pandas DataFrame construction. Avoiding a data copy was actually making it _slower_. Now we simply fill arrays as usual and pass them to the DataFrame constructor.

Fixes performance bug raised in issue 100.

reikdas fixed a case where non-`std::vector` of `TLorentzVector` was not recognized as being readable.

Try again to get `flatten` parameter into PyPI.

Added `flatten=False` as a parameter to all array-fetching methods except LazyArrays, which turns JaggedArrays into regular Numpy arrays by ignoring the starts/stops.

For `outputtype=pandas.DataFrame`, the `flatten=True` case makes the DataFrame's index a `pandas.MultiIndex` with "entry" and "subentry" levels. Branches with different jagged structure can be mixed because `pandas.merge` aligns them appropriately.

Added support for 2D histograms. (Thanks, demarley!)