cumulative deaths from incident deaths

[stephenturner]

I think this fixes the issue of creating cumulative death forecasts from incident deaths, see 5786cac.

High level:

1. Gets the yearweek of the last week of recorded data
2. Gets the first yearweek of forecasted inc deaths
3. Checks that last week of recorded data plus one week equals the first week of forecasted data. If so, proceed.
4. Capture the cumulative deaths recorded on the last week of recorded data in step 1, and hangs on to that for later.
5. Groups by each quantile (or point estimate)
6. Calculates the cumulative sum of incident deaths week over week. So 1 week ahead = 1 week ahead. 2 weeks ahead cumsum = 1+2. 3 weeks ahead = inc deaths for 1+2+3. 4th week = 1+2+3+4.
7. Adds the last week's recorded cumulative deaths from step 4 and adds to the cumsum for each week in step 6.

Ideally this should be a function call, but it's worth thinking through how to safely execute this to guard against supplying data that's unrelated to the forecast you're trying to produce. Turning this into a function as the code stands now would require a user input data (here, the usa tsibble, which contains the ccases week over week), and the forecast (here icases_forecast). You'd want some way to validate that icases_forecast was produced using usa as the input data. The stopifnot should catch egregious errors, but if we want to (over?)engineer this thing to guard against mis-specifying this, we might want to output a list or tie the name of the data used to create the forecast to the attributes or something of the output from your suggested ts_fit() function-in-development.

[vpnagraj]

just ran through everything. made one minor commit to comment out some stray code. will start thinking through how to pull this into the forthcoming ts_fit() ... merging for now