[Combiner] Add a class for computing variance distributedly

[Kelvin-Ng]

Originally in Random Forest in old Husky.

Now it is useful for @Gyingguo 's data visualization project.

[kygx-legend]

I don't quite understand why put this class in combiner. Is it also a combiner or used by combiner？And could you add the unit test, which help users understand?

[Kelvin-Ng]

Very rough pseudo code:

auto ch = create_push_combine_channel<SumCombiner<VarianceMeanNum>>();
VarianceMeanNum a(xx, yy, zz);
for all element x in list:
    a += x;
ch.push(0, a);
ch.flush();
ch.get(0).get_variance() // variance of the whole list

[ddmbr]

This is to calculate variance/standard deviation incrementally. Obviously it's not easy to give a good name...

[kygx-legend]

But I think it should have a unit test for this.

[kygx-legend]

VarianceMeanNum() : variance_(.0), mean_(.0), num_(0) {}