v5.4.0

• Add 4.25.lite model.
• Fix error when sys.stdout is None in some circumstances.
• Improve performance by using separate streams for transfering tensors between CPU and GPU.
• Improve performance by caching Head module output.
• Remove num_streams parameter since it doesn't provide any benefit now due to the above two improvements.
• Re-fix jaggy/wobbling line issue by a significantly faster method than v5.3.0. Using trt=True now requires tensorrt>=10.4.0 and torch_tensorrt>=2.6.0.dev20241011 to fully work. Due to some issues in Torch-TensorRT, dynamic shapes support is temporarily disabled at the moment.