Add TFDWConv() depth_multiplier arg

[glenn-jocher]

Enables grouped non c1 == c2 convolutions in TF YOLOv5 models.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced compatibility for TensorFlow depth-wise convolution layers in YOLOv5 models.

📊 Key Changes

• Adjusted the assertion in the TFDWConv layer to allow the output channels to be a multiple of the input channels, not strictly equal.
• Added depth_multiplier argument to the DepthwiseConv2D layer to handle the case where output channels are multiple of input channels.

🎯 Purpose & Impact

• 🎨 Flexibility: Allows more flexibility in model design by supporting depth-wise convolutions with varied channel proportions.
• 🎓 Compatibility: Ensures broader compatibility with different model architectures that use depth-wise convolutions.
• 🚀 Potential Impact: Users can design and train more complex models within TensorFlow using YOLOv5, potentially improving performance on various tasks.

[glenn-jocher]

Validated