Uproot

Looking up branches (manually by the user or internally in `TTree.arrays`) is now faster for large numbers of branches. The O(n^2) lookup algorithm has been replaced by a single hashmap lookup, and a branch's `interpretation` property is now only computed once.

Three possibilities for `flatten` in `TTree.pandas.df` (and other array-fetching functions with `outputtype=pandas.DataFrame)`):
* `True`: jagged arrays are turned into a `pandas.MultiIndex` and maintain their Numpyness
* `False`: jagged arrays are turned into a column of Python lists
* `None` (new): jagged arrays are excluded, overriding the `branches` parameter.

Note that `flatten=True` is the default for `TTree.pandas.df` and `flatten=False` is the default for other array-fetching methods with `outputtype=pandas.DataFrame`.

This version also fixes an uproot 2 → 3 migration bug in handling jagged arrays.

Adds a `timeout` parameter to XRootDSource, which can be used as following:

python
import uproot
f = uproot.open("root://wherever//whatever.root", xrootdsource={"timeout": 1})


Also, *Source.defaults have been turned into dicts, rather than functions, and dict-based parameters like `xrootdsource` in the above example override defaults, rather than replacing them completely. That way, you don't need to specify the `chunkbytes` and `limitbytes` if you don't want to.

Added support for IPython key completions (152).

Added support for reading files created by GEANT4 (147). The GEANT4 ROOT-file writing engine is independent of ROOT and produces classes with old streamer versions (some don't even have byte counts; see the pull request comments). This allows reading of `tests/samples/from-geant4.root`, which contains TTrees, TH1Ds, and TH2Ds.

Fixed 146. See issue for details.