Uproot

PR 257: consistency in Pandas index for jagged and flat columns.

One quick patch: protect new prefetching mechanism from empty datasets (17923e00eb72ec6728b863700ef44b8138fd3459).

And one near-mistake: homogenized `asarray.destination` return type (662547c33160c3a4ffce28b54d0ed04586f834d8) and then restored it (7e4ad708bd7fcd0d73ab3a203ed8ded5410329c9) because it really needs that different return type.

Fixed 246: error messages pertaining to file formatting now print out which file has failed. Useful if you are running a large job over many files and want to find the one that broke the job.

Added support for reading `std::map<std::string, X>` where `X` is a numerical type or a `std::string`. (PR 245)

Requests for `array(...)` or `arrays(...)` through HTTP and XRootD now start asynchronous downloads of all the basket data before starting to read, decompress, and interpret. This keeps the network busy prefetching while the CPU is preoccupied, hiding latency. The HTTP preloader is implemented with `concurrent.futures.ThreadPoolExecutor` (not selected by default in Python 2, as that would require a non-standard library dependency), and the XRootD preloader is implemented with a pyxrootd callback. The `threads` parameter is a number of threads for HTTP and a boolean for XRootD: yes-parallelize or no-don't, because we don't control how many threads pyxrootd uses. (PR 242)

Binder now uses JupyterLab, rather than Jupyter Notebook. (PR 244)

Moved Event.root to an HTTP source to make the tests directory smaller.

`uproot.numentries` is now documented.