Implement CPU operator for DCNv3

[hxaxd]

Purpose

• Reduce the deployment difficulty of small-scale visual models using DCNv3 on different devices (using the same precompiled package for downgrading)

Work

• Provide CPU downgrade by modifying dcnv3.h
• Modify setup.py to provide CPU-Only compilation method and enable O2 optimization
• Fully implement CPU operators

Effects

• Minimize intrusion into the original code and compilation methods as much as possible
• Accuracy passes tests (based on the original tests by modifying CUDA interfaces to corresponding CPU interfaces)

Issues

• In the scenario of multi-core x86 CPU supporting SIMD, the speed is only 0.27x of the PyTorch CPU version