Mismatched data/benchmark dates creates an error at the end of simulation

[jlowin]

As described in loader.py, benchmark returns are cut off 2 days before the current trading day. But if users run algorithms with a data source that includes the most recent trading days, it creates an error during the risk calculations at the end of the simulation, because the "extra" day will get nan for the benchmark return.

It's happening this weekend: zipline.data.load_from_yahoo() is returning data through 1/21, but the last benchmark return from loader.load_market_data is 1/20.

[freddiev4]

The data is filled through the current trading day (x-ref #2044) so going to close this; we also don't use load_from_yahoo() anymore.