Normalization is inefficient in memory and time

[markotoplak]

• What's wrong?
  Normalization is really slow. I did normalization and PCA on the 100000x1000 data set, and PCA was much faster than normalization. Here is what it looks like (memory use/against time from [ENH] Domain transformations in batches for less memory use #5218):

Orange's normalization, which, by default only needs the mean and standard deviation, computes these very inefficiently. As intermediate result it builds a distribution for each variable, which is, for continuous variables, mostly a sorted list of values.

Because normalization is the default for some learners and unsupervised methods, we should make it faster. This should be huge speedup for some Orange's learners, k-means and PCA.

[janezd]

Check whether we still compute the mean via contintengency table; fix if necessary.

[markotoplak]

How the same graph looks now. Green is current master, red is the old one.