Improved efficiency in the forward part of the decoder.

[tyui592]

https://github.com/Atten4Vis/ConditionalDETR/blob/8ea7c3f3fff49cd287855d4f13cc12a6066bc8e9/models/transformer.py#L145

Since reference_points does not change in the for loop, it is more efficient that pre-calculate the obj_center.