Explore/implement NCV

[Aariq]

I'm more and more convinced that we don't need this, but it might still be good to explore. Copied from usa-npn/cales-thermal-calendars#30

So far NCV (as opposed to REML) seems like it is going to be the way to go. A lot needs to be done to make it work still though, and it's possibly not worth it (or it goes in a separate paper).

• Do comparisons to REML with the end product—marginal slopes
• Create nei object with 3D "neighborhoods" (x, y, and time) to deal with short-range temporal autocorrelation (temporal autocorrelation appears to not be an issue)
• Get correct # of knots

Possibly helpful: Moran's I for estimating spatial autocorrelation of residuals. E.g.

years <- 1981:2023
walk(years, \(.x) {
  gdd_df |>
    mutate(.resid = residuals(m_reml)) |>
    filter(year == .x) |>
    select(x, y, .resid) |>
    rast() |>
    autocor(global = FALSE) |>
    plot(main = paste("Moran's I for", .x))
})

Or just get a single number with ... |> rast() |> autocor(global = TRUE). This could be useful for comparing REML and NCV to check that it actually helps with spatial autocorrelation. Residuals should show no autocorrelation.