* refactor s2anet to support batch_size > 1

* fix problem of inference

* support batch_size > 1 for training

* fix empty results

* fix dota eval

* fix configs of s2anet_head

* modify s2anet_spine_1x to 73 mAP