add support for tengine uint8

[un-knight]

Add support for tengine uint8 backend, mainly refers to:

• https://github.com/OAID/Tengine/blob/45863aa6d928f5967bf05811cada6a166a20ca91/tools/convert_tool/utils/graph_optimizer/graph_opt.cpp#L931
• https://github.com/OAID/Tengine/blob/45863aa6d928f5967bf05811cada6a166a20ca91/tools/quantize/quant_save_graph.cpp#L38
• https://github.com/OAID/Tengine/blob/45863aa6d928f5967bf05811cada6a166a20ca91/tools/quantize/quant_tool_uint8.cpp#L89

[Tracin]

Not compatible with torch v1.81?

[un-knight]

nope, since my developed environment is torch 1.9.0, I remove the requirement of PyTorch 1.8.1 temporarily. But this tengine expansion can be totally compatible with torch v1.8.1, and I will restore the requirements.

[Tracin]

We hightly recommend to add few test cases for this backend !

[un-knight]

We hightly recommend to add few test cases for this backend !

sure, it's necessary to add test cases, but what kind of test cases do we need? Will a table of metrics about imagenet's accuracy with a model before and after quantization be a helper?

[Tracin]

We hightly recommend to add few test cases for this backend !

sure, it's necessary to add test cases, but what kind of test cases do we need? Will a table of metrics about imagenet's accuracy with a model before and after quantization be a helper?

Something like https://github.com/ModelTC/MQBench/blob/main/test/backend/test_backend.py will be enough for now.
Any table of metrics on datasets can be very convincing if that is possible !

[un-knight]

There is MobileNetV2's accuracy on the ImageNet-1k classification task. Since full training of ImageNet is time-consuming, I just finetune the model for several steps, and this is why the top1 accuracy is lower than expected.

model	top1@QAT_tengine_u8	top1@tengine_u8_x86
mobilenet_v2	55.06	54.862

[XHPlus]

Thanks for the awesome PR and looking forward to your continuous contribution! @un-knight