[jvm-packages] buildGroups is unnecessarily expensive

[a-johnston]

Added in #3387, the method relies on a loop followed by two Seq.apply calls:

while (i < groups.length) {
    if (groups(i) != groups(i - 1)) {
...

which for lazily generated sequences (when I checked, groups was an instance of scala.collection.immutable.Stream$Cons) is quadratic. This behavior can be seen by comparing the running times of the following accesses:

val s = Stream.continually(0).toSeq
s(0)
s(1000000)
s(100000000)

(I ctrl-C'd before that last one finished on my machine)

The method buildGroups should do a single pass and not rely on index access.

[CodingCat]

@a-johnston are you interested in filing a PR?

[a-johnston]

Yep, in progress