Rewrite approx (#7214)

This PR rewrites the approx tree method to use codebase from hist for better performance and code sharing.

The rewrite has many benefits:
- Support for both `max_leaves` and `max_depth`.
- Support for `grow_policy`.
- Support for mono constraint.
- Support for feature weights.
- Support for easier bin configuration (`max_bin`).
- Support for categorical data.
- Faster performance for most of the datasets. (many times faster)
- Support for prediction cache.
- Significantly better performance for external memory.
- Unites the code base between approx and hist.

tests/cpp/tree/test_gpu_hist.cu

@@ -315,57 +315,6 @@ TEST(GpuHist, TestHistogramIndex) {
   TestHistogramIndexImpl();
 }
 
-// gamma is an alias of min_split_loss
-int32_t TestMinSplitLoss(DMatrix* dmat, float gamma, HostDeviceVector<GradientPair>* gpair) {
-  Args args {
-    {"max_depth", "1"},
-    {"max_leaves", "0"},
-
-    // Disable all other parameters.
-    {"colsample_bynode", "1"},
-    {"colsample_bylevel", "1"},
-    {"colsample_bytree", "1"},
-    {"min_child_weight", "0.01"},
-    {"reg_alpha", "0"},
-    {"reg_lambda", "0"},
-    {"max_delta_step", "0"},
-
-    // test gamma
-    {"gamma", std::to_string(gamma)}
-  };
-
-  tree::GPUHistMakerSpecialised<GradientPairPrecise> hist_maker{ObjInfo{ObjInfo::kRegression}};
-  GenericParameter generic_param(CreateEmptyGenericParam(0));
-  hist_maker.Configure(args, &generic_param);
-
-  RegTree tree;
-  hist_maker.Update(gpair, dmat, {&tree});
-
-  auto n_nodes = tree.NumExtraNodes();
-  return n_nodes;
-}
-
-TEST(GpuHist, MinSplitLoss) {
-  constexpr size_t kRows = 32;
-  constexpr size_t kCols = 16;
-  constexpr float kSparsity = 0.6;
-  auto dmat = RandomDataGenerator(kRows, kCols, kSparsity).Seed(3).GenerateDMatrix();
-  auto gpair = GenerateRandomGradients(kRows);
-
-  {
-    int32_t n_nodes = TestMinSplitLoss(dmat.get(), 0.01, &gpair);
-    // This is not strictly verified, meaning the numeber `2` is whatever GPU_Hist retured
-    // when writing this test, and only used for testing larger gamma (below) does prevent
-    // building tree.
-    ASSERT_EQ(n_nodes, 2);
-  }
-  {
-    int32_t n_nodes = TestMinSplitLoss(dmat.get(), 100.0, &gpair);
-    // No new nodes with gamma == 100.
-    ASSERT_EQ(n_nodes, static_cast<decltype(n_nodes)>(0));
-  }
-}
-
 void UpdateTree(HostDeviceVector<GradientPair>* gpair, DMatrix* dmat,
                 size_t gpu_page_size, RegTree* tree,
                 HostDeviceVector<bst_float>* preds, float subsample = 1.0f,