functions to make it easy to configure image scanners that use the hashed_feature_image.

relative to the number of times they occur in an image.  I also set
the new default behavior to use this relative weighting and changed
the serialization format to accommodate this.

feature extractors can have their parameters queried.

boxes is detected.

precision values to be slightly wrong.

certain rare cases.

instead of mean average precision.

correlation() methods to output the wrong answer if the covariance was
negative.

the order of function argument evaluation is not defined, the test had slightly
different results on different systems which could cause the tests to falsely
indicate failure.

solver.  This makes some things, such as the structural_object_detection_trainer,
significantly faster.

non-negative weight learning support.