redundant inference in the sliding window of test_dota

[lizeshengh]

In the sliding window inference of test_dota, the loops do not stop when the sliding window reach the boundary of the original image. It causes extra inference and slow down the performance. For example, an 800x800 image with h_len=w_len=400 and h_overlap=w_overlap=200 only need 9 sub images for inference. However, in the current code, it loops and infers 16 times.

[lizeshengh]

Changing the stop conditions of those two for-loops in the function "worker()", or using a set to record (hh_, ww_) pairs that has been used can solve the problem.