Forecast Model Builder - fix: handle partial days, weeks, months when we roll up

[sfc-gh-kmason]

For example if rolling up to week level, and one of the weeks has less than 7 days, then that week's values will appear lower than it actually should be. So we need to handle roll-ups that don't contain a full cycle.

This also affects how the VALIDATION_DAYS parameter works. Because it is set to days (instead of time periods), if you roll up to the week (or month) level, then the last week or month might be truncated if the number of days we set aside doesn't match up with the end of a week (or month). Need to check the logic here, and make sure nothing is getting cut off with the train/test split.