Now input vector with class is required for unsupervised classifier.
The cases where input vector data were optional for unsupervised
classification are commented.

Update license for added/moved files.

shark k-means do not provide confidence map (for now).

This allow to have only one application to handle bot supervised and
unsupervised classifier.

Classifier category is retrieve with selected classifier and depending
this category some operation are added/removed for the training.

This workaround is added to avoid a crash in application when build
is made with RelWithDebInfo which is caused by a
boost::archive_exception impossible to be catch.
This workaround break the compatibility with shark model direct
read/write.
It is still possible to read old RF model from OTB.

SampleSelection can now be used without input vector data.

When only image is provided to the application a default vector
corresponding to the image envelope is generated.

All Supervised and Unsupervised Classifier should be declared and stay
in Supervised/MachinLearningModelFactory since multiple application
depend on it.

Use inheritance with otbTrainImagesBase to create unsupervised
composite application.

Create a common class with template to perform Supervised classification
(same as otbTrainImagesClassifier) or Unsupervised.

Add the ability to specify Supervised or Unsupervised classification