Forest serialization

[jemus42]

Implements #52

fit <- rpf(mpg ~ ., data = mtcars)
blob <- rpf_marshal(fit)          # plain R list, saveRDS-safe
saveRDS(blob, "model.rds")
restored <- rpf_unmarshal(readRDS("model.rds"))
predict(restored, mtcars)         # bit-identical to original

Hard requirement to get the mlr3extralearner back.
Unfortunately not entirely trivial and a few scenarios are covered:

Mode	Usage	Blob contains	After restore
Prediction-only (default)	blob <- rpf_marshal(fit)	leaves + bounds	predict() works; purification permanently unavailable
Purified lean (recommended)	purify(fit); blob <- rpf_marshal(fit)	leaves + bounds + purified grids	purified predict() and predict_components() work. No training data stored, no re-purification on load
Purify-later (opt-in)	blob <- rpf_marshal(fit, include_data = TRUE)	leaves + bounds (+ purified grids if purified) + training X/Y	everything, including purify() after restore

In particular I wanted the "don't care about purification" path because that's important for settings like the benchmark(s) we did in the past where only prediction is relevant and serialization is potentially needed (for encapsulation with callr as done in my large survival benchmark, or maybe parallelization across mirai or something along those lines). in that setting forcing purification would be wasted time, same as carrying around the information needed for purification later.