Add training window option to arx_forecaster

[brookslogan]

As @dshemetov and I were comparing some production forecasters to epipredict canned forecasters, it seemed like it would be useful to have a training window as a parameter (implemented via step_training_window).

Some asides:

• Does step_training_window need to be step_epi_training_window, since it relies on epi_keys and time_value?
• I've been using "epikey" (or "epigroup") to refer to the geographical&demographical keys, not including time_value or version, but I think epipredict considers time_value an epikey. We should think about standardizing names.

[brookslogan]

Alternatively: add a function to output the unfit workflow that would be used by arx_forecaster(), have utilities to modify the workflow more easily ([add_]step*(.after = "<step/phasename>", .before = "<step/phasename>", ....)?