vvas_xinfer runs models that mix tensor precisions across layers - for example,
an INT8 backbone with an FP16 detection head. Supported dtypes are INT8,
BF16, FP16 and FLOAT32, on both the ONNX Runtime (VitisAI EP) and VART
backends. The plug-in handles the precision boundaries transparently; no extra
configuration is required beyond pointing vart-model-path (or onnx-model-path)
at the compiled mixed-precision model. Refer to Vitis-AI compiler guide to generate models.