Update GradientBoost to xgboost >= 2.0 API (xgb.DMatrix)

[instantepiphany]

Summary

GradientBoost() was written against the pre-2.0 xgboost API (raw matrix + label = arg). Since xgboost 2.0, xgb.cv() and xgb.train() enforce inherits(data, \"xgb.DMatrix\"), and xgb.cv() no longer accepts label. The legacy xgboost() wrapper also renamed data → x and label → y. As of the version currently shipping in NixR (xgboost 3.2.0.1), every Gradient Boosting test in the regression suite fails before the first round of training with:

```
Error: inherits(data, "xgb.DMatrix") is not TRUE
```

(The error is emitted by `stopifnot()` at xgboost `R/xgb.cv.R:118`, which quotes the unevaluated expression — i.e. the variable is named `data`, not `dada`.)

Surfaced by the Standard R Tests / Machine Learning - Gradient Boosting.Q regression run #7389: log.htm. The Q file's `tree` RItem (and all sibling `xgb*`, `importance tree`, `gradient.boost`, `imputation`, `tree with grid` items) all funnel through `GradientBoost()`, so they all fail the same way.

Changes

• Build `xgb.DMatrix(data = numeric.data$X, label = numeric.data$y)` once and reuse it.
• `xgb.cv()` calls now pass `data = dtrain` with no `label =`. Inside the grid-search worker, `set.seed(seed)` is called explicitly before each `xgb.cv` (the old `seed =` parameter was dropped from `xgb.cv`).
• Replace `xgboost(data = …, label = …, params = …, nrounds = …)` with `xgb.train(data = dtrain, params = …, nrounds = …)` so the returned object is still an `xgb.Booster` and the downstream contract on `result$original` (`$params`, `$call`, `$feature_names`, `$raw`, `$handle`) is unchanged.
• Read `xgbcv$early_stop$best_iteration` instead of `xgbcv$best_iteration` (relocated in xgb.cv's return shape).
• `@importFrom` / NAMESPACE: drop `xgboost`, add `xgb.DMatrix` and `xgb.train`.
• DESCRIPTION: bump version to 1.2.5 and require `xgboost (>= 2.0.0)`.

Reproduction

Reproduce the original failure (raw matrix into modern xgb.cv):

```bash
cd /path/to/r-server
nix run .#R -- --vanilla -e 'set.seed(1)
x <- matrix(rnorm(40), 20, 2)
y <- sample(0:1, 20, replace = TRUE)
xgboost::xgb.cv(data = x, label = y, nrounds = 2, nfold = 3,
params = list(objective = "binary:logistic"), verbose = 0)'

-> Error: inherits(data, "xgb.DMatrix") is not TRUE

-> Warning: Parameter(s) have been removed from this function: label.

```

Confirm the new code path works against the same xgboost:

```bash
nix run .#R -- --vanilla -e 'library(xgboost)
set.seed(1)
x <- matrix(rnorm(60), 30, 2)
y <- sample(0:1, 30, replace = TRUE)
dtrain <- xgb.DMatrix(data = x, label = y)
xgb.cv(data = dtrain, nrounds = 5, nfold = 3,
params = list(objective = "binary:logistic", num_class = 1, max_depth = 3),
verbose = 0)
xgb.train(data = dtrain,
params = list(objective = "binary:logistic", num_class = 1, max_depth = 3),
nrounds = 3, verbose = 0)'

-> xgb.cv ok; xgb.train returns class "xgb.Booster"

```

Build this branch as an installable R package via the NixR adhoc-deployment branch (override the flipMultivariates input to point at your local checkout of this branch):

```bash
nix build 'git+ssh://git@github.com/displayr/NixR?ref=untracked-update-nixpkgs#pkgs.x86_64-linux.rPackages.flipMultivariates' \
--override-input flipMultivariates .
```

Test plan

• `Standard R Tests / Machine Learning - Gradient Boosting.Q` passes against an R server using this branch (re-run of the failing job above).
• Spot-check both grid-search and non-grid-search code paths (`tree` vs `tree with grid` RItems in the same Q file).
• `Importance` output still renders (`importance tree` RItem) — exercises `xgb.importance(model = x$original)`.
• `predict.GradientBoost` smoke test — exercises `x$original$params$objective`, `x$original$feature_names`, and the `xgb.load.raw(x$original$raw)` salvage path in `R/variables.R`.

🤖 Generated with Claude Code