Move prediction cache to Learner. (#5220)

* Move prediction cache into Learner.

* Clean-ups

- Remove duplicated cache in Learner and GBM.
- Remove ad-hoc fix of invalid cache.
- Remove `PredictFromCache` in predictors.
- Remove prediction cache for linear altogether, as it's only moving the
  prediction into training process but doesn't provide any actual overall speed
  gain.
- The cache is now unique to Learner, which means the ownership is no longer
  shared by any other components.

* Changes

- Add version to prediction cache.
- Use weak ptr to check expired DMatrix.
- Pass shared pointer instead of raw pointer.

tests/cpp/helpers.h

@@ -16,16 +16,13 @@
 
 #include <dmlc/filesystem.h>
 #include <xgboost/base.h>
-#include <xgboost/objective.h>
-#include <xgboost/metric.h>
 #include <xgboost/json.h>
-#include <xgboost/predictor.h>
 #include <xgboost/generic_parameters.h>
 #include <xgboost/c_api.h>
-#include <xgboost/learner.h>
 
 #include "../../src/common/common.h"
 #include "../../src/common/hist_util.h"
+#include "../../src/gbm/gbtree_model.h"
 #if defined(__CUDACC__)
 #include "../../src/data/ellpack_page.cuh"
 #endif
@@ -42,6 +39,12 @@
 #define GPUIDX -1
 #endif
 
+namespace xgboost {
+class ObjFunction;
+class Metric;
+struct LearnerModelParam;
+}
+
 bool FileExists(const std::string& filename);
 
 int64_t GetFileSize(const std::string& filename);
@@ -206,7 +209,7 @@ std::unique_ptr<DMatrix> CreateSparsePageDMatrixWithRC(
     size_t n_rows, size_t n_cols, size_t page_size, bool deterministic,
     const dmlc::TemporaryDirectory& tempdir = dmlc::TemporaryDirectory());
 
-gbm::GBTreeModel CreateTestModel(LearnerModelParam const* param);
+gbm::GBTreeModel CreateTestModel(LearnerModelParam const* param, size_t n_classes = 1);
 
 std::unique_ptr<GradientBooster> CreateTrainedGBM(
     std::string name, Args kwargs, size_t kRows, size_t kCols,