Forest serialization and deserialization

[jemus42]

It needs to be possible to safely store an rpf for practical applications where refitting ad-hoc is not feasible, i.e. the following needs to run without error:

library(randomPlantedForest)

# Normal use
rpfit <- rpf(mpg ~ wt + cyl, data = mtcars)
rpfit
#> -- Regression Random Planted Forest --
#> 
#> Formula: mpg ~ wt + cyl 
#> Fit using 2 predictors and main effects only.
#> Forest is _not_ purified!
#> 
#> Called with parameters:
#> 
#>             loss: L2
#>           ntrees: 50
#>  max_interaction: 1
#>           splits: 30
#>        split_try: 10
#>            t_try: 0.4
#>            delta: 0
#>          epsilon: 0.1
#>    deterministic: FALSE
#>         nthreads: 1
#>           purify: FALSE
#>               cv: FALSE
predict(rpfit, mtcars)
#> # A tibble: 32 × 1
#>    .pred
#>    <dbl>
#>  1  20.4
#>  2  20.3
#>  3  26.1
#>  4  20.8
#>  5  16.5
#>  6  18.6
#>  7  15.1
#>  8  23.8
#>  9  23.1
#> 10  19.0
#> # ℹ 22 more rows

# Serialize and cleanup
temp_loc <- tempfile()
saveRDS(rpfit, file = temp_loc)
rm(rpfit)

# Attempt to restore: Not working
rpfit <- readRDS(temp_loc)
rpfit
#> -- Regression Random Planted Forest --
#> 
#> Formula: mpg ~ wt + cyl 
#> Fit using 2 predictors and main effects only.
#> Error in .External(structure(list(name = "CppMethod__invoke_notvoid", : NULL value passed as symbol address
predict(rpfit, mtcars)
#> Error in .External(structure(list(name = "CppMethod__invoke_notvoid", : NULL value passed as symbol address

^{Created on 2023-11-21 with reprex v2.0.2}

As of now I don't know how the internals work for that, or if it's even possible to hook into saveRDS for that or if we need to write a separate serialization mechanism.

It's not a high priority now, but in the long run this needs to be possible.

[jemus42]

#55 contains the necessary machinery in src/lib/rpf.cpp to reconstruct a given rpf model, which together with the get_model() internal method should already be mostly sufficient to solve this.

As far as I can tell there's no easy way to hook into saveRDS or anything, so I'm thinking a simple save_rpf() / load_rpf() type API should do the trick for now.

In any case I'll have to make sure the reconstructed object also contains all the extra bits that the object returned by rpf() stores, but that should not be a big issue.

In addition, I thought it would be nice to also add bundle S3 methods, so we can directly hook into an existing attempt to standardize the whole model serialization shebang.
The xgboost method should suffice as a point of orientation, and bundle would then be a Suggest: I guess?
I know there's a way to provide a method for a generic you don't want to Imports: but I'll have to look that up when the time comes.