normalization constant was being computed incorrectly.

a new function called sum_filter().

calling get_best_matching_rect().

reconstruct them for each separation oracle call.

Also improved its spec.

given the recent changes to the object detection tools.

a scan_image_pyramid when it is used with a hashed_feature_image
and projection_function.

since they logically modify the state of the object detector now that
we are exposing the state of the underlying image scanner.

for the test_box_overlap object when it is used and the user
hasn't specifically assigned an overlap tester.

matching with a truth rectangle as well as overlap in the context of
non-max suppression of results.  Since these are two very separate
ideas I renamed some things to avoid any confusion.

by a template argument rather than its own method.  By default,
this object will use the projection_hash.

get_feature_vector() method into two separate functions so the interface
is a little simpler and more flexible.