[SPARK-5365][MLlib] Refactor KMeans to reduce redundant data

[viirya]

If a point is selected as new centers for many runs, it would collect many redundant data. This pr refactors it.

[srowen]

So this returns (p, (r1, r2, r3, ...)) instead of (r1, p), (r2, p), (r3, p), ... Makes sense to me, especially if you have reason to believe this is a bottleneck somewhere.

[viirya]

Especially when there are many runs to use and p is also high dimensional and selected in more than one run. Then collecting redundant p would be too useless and time-consuming.

[SparkQA]

Test build #25962 has finished for PR 4159 at commit 25487e6.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[mengxr]

LGTM. Merged into master. Thanks!