Add PCA-based codebook initialization.

[exaexa]

This initializes the codebook as a simple, straight PCA-oriented 2-dimensional grid.

The involved changes include a slight update of the initialization logic (initialization functions return point data instead of indices). New usage:

• BuildSOM( ... no arguments ...) uses the random initialization
• BuildSOM(init=T) uses the original Initialize function
• BuildSOM(init=T, initf=Initialize_KWSP) is the more verbose default-less version of the previous
• BuildSOM(init=T, initf=Initialize_PCA) is the PCA-based initialization

There are a few tunables involved in the process which would probably deserve extra attention:

• the sdev_scale default value needs to get tested on more data (but overall it works pretty good for me even in this simple form)
• PCA-based initialization basically simulates the "unfolding" of the SOM, which sometimes takes first 20% of the process. It is therefore possible to lower rlen and the initial threshold parameter settings if using this initialization.
• Replacing the initial part of high-plasticity high-alpha SOM training can make the training very predictable and repeatable; but testing on more datasets is needed to see whether anything isn't negatively affected by the `simplification'. I did not see any problems so far, though.

Thanks a lot! -mk