use nms in the model itself, simplify and make postprocess faster.

make inference in batches, fix c# handling, add overlap handling

Azaion.Inference/README.md

@@ -13,6 +13,17 @@ Results (file or annotations) is putted to the other queue, or the same socket,
 
 <h2>Installation</h2>
 
+Prepare correct onnx model from YOLO:
+```python
+from ultralytics import YOLO
+import netron
+
+model = YOLO("azaion.pt")
+model.export(format="onnx", imgsz=1280, nms=True, batch=4)
+netron.start('azaion.onnx')
+```    
+Read carefully about [export arguments](https://docs.ultralytics.com/modes/export/), you have to use nms=True, and batching with a proper batch size  
+
 <h3>Install libs</h3>
 https://www.python.org/downloads/
 
@@ -45,7 +56,7 @@ This is crucial for the build because build needs Python.h header and other file
 
 ```
     python -m pip install --upgrade pip
-    pip install opencv-python cython msgpack cryptography rstream pika zmq pyjwt pyinstaller tensorboard
+    pip install requirements.txt
 ```
 In case of  fbgemm.dll  error (Windows specific):