Remove CStatistic::mean/variance/covariance

[karlnapf]

And replace all their usage with linalg implementation. In addition, respect Shoguns way of defining vectors, which is as column vectors, not row vectors. A couple of unit tests and other places in Shogun depend on this.

Mean/Std/Var:

• partly already in linalg
• should be possible to compute elementwise, rowwise, and columnwise

Covariance:

• not in linalg yet, but should re-use mean computation

See also #3117

[c4goldsw]

@karlnapf I'm trying to find a standard deviation or variance function in linalg, but I'm unable to find one - could you please point me in the right direction? Or, have I misunderstood you, in that it has yet to be implemented (because of "partly already in linalg")?

[c4goldsw]

Also, I've noticed that there are different implementations of functions for vectors and matrices - I've replaced all usages of CStatistics::mean() with linalg::mean() (#3096), but I haven't done anything with CStatistics::matrix_mean() . Does this issue involving replacing both or only vector implementations (obiviously for covariance we're just working with matricies)?

[karlnapf]

The idea is to

1. Add all functionality for mean/variance/etc to linalg. Those that are not yet there, such as variance
2. Make Shogun use them rather than CStatistics
3. Remove them from CStatistics

Start one by one. As a first step, you could grep for CStatistics::matrix_mean etc and replace them with linalg calls