ppnwp: Post-Process Numerical Weather Prediction ensembles

Implements different strategies based on on sliding/time-of-day training and rolling/non-rolling formats. For non-rolling forecasts, sliding training generates only a single model at the issue time using the sequence of data up until one step before issue. Time-of-day training generates unique models for each step in the run, matching the observations from previous days to the forecast issued at the same total lookahead time (lead time + partial horizon).

1 2	get_training_subsets(time_idx_forecast, issue, step, metadata, ensemble, telemetry)

`time_idx_forecast`	Index of forecast time, relative to telemetry's valid times
`step`	Step (index) in this forecast run
`metadata`	A data.frame of forecast parameters
`ensemble`	A list of data=[issue x step x member] array of all ensemble data (historical + test) and issuetime=vector of POSIXct time stamps
`telemetry`	A list of data=vector of telemetry and validtime=vector of POSIXct times