efficiently, even when their window sizes differ.

non-max suppression code that I introduced a few commits ago.

into them.

That is, chain the = operator when using a literal assignment on
matrix objects.

going to be performance critical and is a common user error.

cross_validate_trainer_threaded() routines.

that closesocket() fails.

dense vectors when assigning them to sparse vectors.  Also updated dot()
so that it is much faster when working with "sparse vectors" that are
the product of assigning a dense vector to a sparse vector.

svm_multiclass_linear_trainer.