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/gbm/test_gbtree.cc

@@ -10,6 +10,7 @@
 #include "xgboost/learner.h"
 #include "../helpers.h"
 #include "../../../src/gbm/gbtree.h"
+#include "xgboost/predictor.h"
 
 namespace xgboost {
 TEST(GBTree, SelectTreeMethod) {
@@ -22,9 +23,8 @@ TEST(GBTree, SelectTreeMethod) {
   mparam.num_feature = kCols;
   mparam.num_output_group = 1;
 
-  std::vector<std::shared_ptr<DMatrix> > caches;
   std::unique_ptr<GradientBooster> p_gbm {
-    GradientBooster::Create("gbtree", &generic_param, &mparam, caches)};
+    GradientBooster::Create("gbtree", &generic_param, &mparam)};
   auto& gbtree = dynamic_cast<gbm::GBTree&> (*p_gbm);
 
   // Test if `tree_method` can be set