use nms in the model itself, simplify and make postprocess faster.

make inference in batches, fix c# handling, add overlap handling

Azaion.Inference/annotation.pxd

@@ -1,10 +1,17 @@
 cdef class Detection:
     cdef public double x, y, w, h, confidence
+    cdef public str annotation_name
     cdef public int cls
 
+    cdef public overlaps(self, Detection det2)
+
 cdef class Annotation:
-    cdef bytes image
+    cdef public str name
+    cdef public str original_media_name
     cdef long time
     cdef public list[Detection] detections
+    cdef public bytes image
+
+    cdef format_time(self, ms)
     cdef bytes serialize(self)