Consider branching targets over forecast_date

[dshemetov]

We do this in the prod pipelines, but not explore. This has a few benefits:

• built-in per date caching, so redoing backtests on a slightly different date range is faster
• removes the caching hack we do with epix_slide_simple (see the bottom of the joined_archive_data target)
• removes the need for epix_slide_simple entirely

I tried extending this to explore and ran into a few issues:

• I made nested tar_map calls, first over forecast_date and then over forecasters and their parameters; this worked fine
• however, I got tangled in the scoring; I now realize that we can simply score each forecaster-date target and then aggregate after

Branching like this should save a lot of time in exploration, since adding a single date of forecasting will not require rerunning a 3 hour pipeline.