CUDA ops

We implement common CUDA ops used in detection, segmentation, etc.

• BBoxOverlaps

• CARAFE

• CrissCrossAttention

• ContextBlock

• CornerPool

• Deformable Convolution v1/v2

• Deformable RoIPool

• GeneralizedAttention

• MaskedConv

• NMS

• PSAMask

• RoIPool

• RoIAlign

• SimpleRoIAlign

• SigmoidFocalLoss

• SoftmaxFocalLoss

• SoftNMS

• Synchronized BatchNorm

• Weight standardization