Update precision/recall after subsetting of confusion matrix

[thinkh]

Currently the metrics curves (precision, recall, F1 score) are always computed for the whole confusion matrix. When filtering classes (i.e., subsetting the confusion matrix) these curves must be updated based on the selected set of classes.

Performance metrics for 10 classes:

Same performance metrics for 4 classes:

@gfrogat Check how the subsetting/computation is handled.

[thinkh]

It seems that it is already correct, please check the computed performance values.