* sync TrainReader.batch to global train_batch_size

* Add `matrix_nms` op

* Make OP registration more robust

* Polish Reader to simplify preprocessing logic.
#  sample_transforms-> make batch -> batch_transforms in Reader.
* Clean some code
* Imporve yolov3_r50vd_dcn_obj365_pretrained_coco 41.4 to 41.8.
* Update all configs.

* add multi-scale test

* Clean up a bit

* Add `__shared__` annotation for shared variables between modules

* Dedup `num_classes` configurations

* Document `__shared__` usage

* Remove some unused variable

* Improve docstring and add comments for `__shared__`

some of the global config values are reference type, e.g., objects, lists or
dicts, if the created module instances modify those values (see FPN and
spatial_scale), global config will reflect these changes, and instances of the
same class created later will inherit the changed values

* Rename object_detection to PaddleDetection
* Small fix for doc