The CSV format is much more compact (does not repeat field names for
every row), and more naturally fits with R anyway.

Alter the relevant tests to serve CSVs. I've verified all vignettes
build with these changes.

It should not be possible to have two signals with the same source,
signal, time_type, and geo_type. This will cause a query for that signal
to have two metadata rows attached to the covidcast_signal data frame,
which will confuse everything.

Fetching multiple days is important.

Profiling revealed that latest_issue was responsible for a large portion
of the time taken in building correlation-utils.Rmd (apart from
downloading the data). Much of this time was spent in dplyr::filter.

Rather than grouping by geography and time, we can use dplyr::distinct,
knowing that each geo_value and time_value should appear only once per
issue date. By taking the first or last (after sorting by issue date),
we get the desired result.

dplyr does not document algorithmic details, so I can't easily give O(n)
notation here. Algorithmic details notwithstanding, the results are
extraordinary:

> nrow(d)
[1] 203360
> system.time(latest_issue_old(d))
   user  system elapsed
  6.395   0.037   6.465
> system.time(latest_issue(d))
   user  system elapsed
  0.025   0.003   0.027

Fetching the county data took a large portion of the time required to
build the vignette, particularly after the fixes to latest_issue in
b0f7e7b.

Always run your tests before pushing...

By providing the `repo` block with a link pointing to
R-packages/covidcast/, pkgdown can build the correct URLs.

Also modify several vignettes to download only the necessary amount of
data, thus reducing the file size of these CSV files.