[core] fix slow predict-caching with many classes

[khotilov]

Addresses the issue of the O(num_class^2) prediction cache behavior #1689, #2926 where cache updates within CommitModel may start dominating with many classes. The reason was that the cache was updated for each single class group commit in a boosting round. E.g., with

set.seed(1)
n <- 1e4
num_feat <- 200
num_class <- 150
y <- apply(rmultinom(n, 1, rep(1, num_class)), 2, function(yy) which(yy != 0)) - 1
dtr <- xgb.DMatrix(matrix(rnorm(n*num_feat), n, num_feat), label = y)
param <- list(objective='multi:softprob', num_class=num_class, debug_verbose=1,
              tree_method='hist', subsample=0.6)
bst <- xgb.train(param, data=dtr, nrounds=1)
rm(bst)
gc()

the timing result before (note that no watchlist was used, and using it would make caching even slower):

[16:21:50] ======== Monitor: GBTree ========
[16:21:50] BoostNewTrees:	 6.436368s
[16:21:50] CommitModel:	 10.707612s

and after the fix:

[16:22:51] ======== Monitor: GBTree ========
[16:22:51] BoostNewTrees:	 6.497372s
[16:22:51] CommitModel:	 0.109006s

Some minor changes:

• remove redundant 'if' in cpu_predictor
• get rid of compiler warnings
• make R on windows not to complain about configure script by providing an empty configure.win
• workaround for R v3.4.3 bug Error trying to compile with GPU support for R #3081

[codecov-io]

Codecov Report

Merging #3109 into master will decrease coverage by <.01%.
The diff coverage is 10.52%.

@@             Coverage Diff              @@
##             master    #3109      +/-   ##
============================================
- Coverage     43.79%   43.79%   -0.01%     
  Complexity      228      228              
============================================
  Files           159      159              
  Lines         12507    12507              
  Branches        466      466              
============================================
- Hits           5478     5477       -1     
- Misses         6837     6838       +1     
  Partials        192      192

Impacted Files	Coverage Δ	Complexity Δ
src/gbm/gbtree.cc	17.95% <0%> (-0.1%)	0 <0> (ø)
src/objective/regression_obj.cc	84% <100%> (ø)	0 <0> (ø)	⬇️
src/predictor/cpu_predictor.cc	68.71% <50%> (+0.19%)	0 <0> (ø)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 375d753...c414b00. Read the comment docs.