[WIP][SPARK-19208][ML][MLLib] Optimize MultivariantOnlineSummerizer by making the summarized target configurable

[WeichenXu123]

What changes were proposed in this pull request?

add a mask parameter to MultivariantOnlineSummerizer constructor.
it can be the following values now:
meanMask
varianceMask
minMask
maxMask
numNonZerosMask

so that we can config the summarized targets in the following way:
new MultivariantOnlineSummerizer(meanMask|varianceMask)
it represent this summarizer will only compute mean and variance.

How was this patch tested?

unit test added.

[SparkQA]

Test build #64902 has finished for PR 14950 at commit dc44bb9.

• This patch fails Scala style tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
  • class MultivariateOnlineSummarizer(mask: Int)

[srowen]

Hm how much does this really save? these are pretty cheap sufficient statistics. Now you have to know whether your particular object was configured or not to return the answer you want.

[SparkQA]

Test build #64904 has finished for PR 14950 at commit be286eb.

• This patch fails Scala style tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
  • class MultivariateOnlineSummarizer(mask: Int)

[WeichenXu123]

@srowen not only cpu cost, if data dimension is big, serialization cost will be big, such as #14109
and compute all target seems not proper if we may add more summary targets in the future ?

[SparkQA]

Test build #64905 has finished for PR 14950 at commit 3029468.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
  • class MultivariateOnlineSummarizer(mask: Int)

[WeichenXu123]

when benchmark is done I will reopen it.