feat(*): add neuron adapter api as draft

[lgyStoic]

add neuro backend as draft

[CLAassistant]

All committers have signed the CLA.

[lgyStoic]

https://pytorch.org/executorch/stable/build-run-mediatek-backend.html as reference, I got sdk and mtk adapter header from executorch repo, and mtk adapter api is quit similar as nnapi interface. @jxt1234

[lgyStoic]

@wangzhaode @jxt1234
Thus far, we have successfully implemented basic support for all operators in ResNet101 (such as convolution, addition, raster, etc.), demonstrating that this approach is indeed viable. Moreover, on the Dimensity 9000 and 9200 platforms, the speed significantly surpasses that of NNAPI. However, several issues remain to be addressed:

• The MTK Adapter API closely resembles NNAPI but exhibits certain distinctions. If these differences prevent code reuse, it will result in the existence of two nearly identical codebases, complicating maintenance efforts.
• It is imperative to complete the implementation of the remaining operators to ensure compatibility with a broader range of widely used neural networks.
• Introducing a caching mechanism is essential to eliminate the need for recompilation with each execution, a feature that NNAPI also incorporates.

simple benchmark blow

[github-actions]

Marking as stale. No activity in 60 days.

[FranzKafkaYu]

So can we use this PR for MTK APU acceleration in ResNet model for now?

[lgyStoic]

So can we use this PR for MTK APU acceleration in ResNet model for now?

@FranzKafkaYu yep，this pull request is already done all dependency for running restnet on mtk neuron pilot; also can run quantized int8 mnn model on neuron plot; demo upper show that ablity, for simpliy I use reset101 model work when click mobilenet button, and show correct classification result.