PipeOpFeatureUnion should use DataBackend cbind

[mb706]

PipeOpFeatureUnion currently uses Task$data() to get the data to cbind. We can save some space and performance if we just cbind the DataBackend directly.

Difficulty: The DataBackends do not contain any information about col roles (in particular which columns removed from the data), so this has to be done after cbind intelligently.

[sumny]

So how much of improvement can we expect here. I recently changed the cbind_tasks method of PipeOpFeatureUnion to rely on modifications by reference via data.table where possible (so there are only the $data() calls in line 179 left which could be bad memorywise/speedwise). However, on my machine the current changes #358 already yield drastic improvements for "larger" data:

dat1 = matrix(rnorm(1000 * 100000), nrow = 100000, ncol = 1000)
colnames(dat1) = c("target", paste0("feature", 1:999))
dat1 = as.data.table(dat1)
task1 = TaskRegr$new("test1", backend = dat1, target = "target")

dat2 = cbind(dat1[, 1], matrix(rnorm(999 * 100000), nrow = 100000, ncol = 999))
colnames(dat2) = c("target", paste0("feature", 1000:1998))
dat2 = as.data.table(dat2)
task2 = TaskRegr$new("test2", backend = dat2, target = "target")

po = PipeOpFeatureUnion$new()
microbenchmark::microbenchmark(po$train(list(task1, task2)), times = 10L)

#!!!Unit: milliseconds
#                         expr      min       lq     mean   median      uq
# po$train(list(task1, task2)) 452.4882 543.9716 640.1103 601.5686 685.713
#      max neval
# 1046.604    10

vs. current master

#!!!Unit: seconds
#                         expr      min       lq     mean  median       uq
# po$train(list(task1, task2)) 191.6633 212.4702 215.7789 215.179 218.7174
#      max neval
# 243.1069    10

which is a median improvement of around the factor ~ 350.

[mb706]

It would still be useful where we are using non-data.table backends. But it would also be a bit difficult, so this has low priority for now.