Step through pre-calculated start times for each group using closure rather than using .real col in epi_slide

[nmdefries]

Instead of re-calculating .ref_time_values, use pre-calculated values stored in starts. For each group, use a closure to keep track of position in starts vector. This avoids the slow .real filtering and removal steps, and simplifies the code by removing various bits of .real handling.

This is ~4x faster than the old version.

[nmdefries]

The test failures appear to be a problem with the new withr (3.0.0, released yesterday); tests pass locally.

[nmdefries]

All the man/*.Rd files are from styling changes.

[brookslogan]

Looks good, please see minor TODOs above.

Nice 4x!!! Was this on something like 7-day averaging? [I'm getting something like a 2x doing jhu_csse_daily_subset %>% group_by(geo_value) %>% epi_slide(before = 6, cases_7davish = mean(cases)), 3x on an ungrouped version. Still super nice.]

[nmdefries]

My (main) test case was indeed a 7-dav, calling the exploration-tooling rolling mean fn on a synthetic dataset,

n_days <- 4000
removed_date <- 10
simple_dates <- seq(as.Date("2012-01-01"), by = "day", length.out = n_days)
simple_dates <- simple_dates[-removed_date]
rand_vals <- rnorm(n_days - 1)

# Three states, with 2 variables. a is linear, going up in one state and down in the other
# b is just random
# note that day 10 is missing
epi_data <- epiprocess::as_epi_df(rbind(tibble(
  geo_value = "al",
  time_value = simple_dates,
  a = 1:(n_days - 1),
  b = rand_vals
), tibble(
  geo_value = "ca",
  time_value = simple_dates,
  a = (n_days - 1):1,
  b = rand_vals + 10
), tibble(
  geo_value = "fl",
  time_value = simple_dates,
  a = (n_days - 1):1,
  b = rand_vals * 2
)))

Surprisingly,

jhu_csse_daily_subset %>% group_by(geo_value) %>% epi_slide(before = 6, ~mean(.x$cases), new_col_name = "cases_7davish")

is twice as fast as

jhu_csse_daily_subset %>% group_by(geo_value) %>% epi_slide(before = 6, cases_7davish = mean(cases))

It looks like converting a quosure to fn is the bottleneck, with this line being particularly slow.