Implement wasi-nn using nntrainer

[kulcsaradam]

Introduce nntrainer as a third party application but do not include it as a submodule. Change build files, implement functions required for wasi-nn functionality and add simple test for wasi-nn.

[kulcsaradam]

The patch changes quite a bit about how files work, since wasi treats files and directories the same, I've merged the two classes. These changes could be done in a different patch if needed.

[kulcsaradam]

The patch will be broken up into a series of smaller patches for manageability's sake. After those patches are merged I'll change the draft to be ready for review.

[kulcsaradam]

In the meantime I would welcome input on the patch, especially on some aspects:

• There is a hack with std as we use c++11 and the nntrainer api is on c++14. It uses make_unique and enable_if_t that are c++14 features. I've implemented a hack that compiles and runs fine in every compiler as I've read (gcc, clang and mscv) but according to the c++ standard is UB. Would this be okay or should we search for another solution?
• NNTrainer is compiled as a shared object file and then loaded by either the system or through preloading. The project is not included in walrus for manageability's sake. I hope this solution is sufficient enough.
• Currently we support nntrainer's .ini files and binary format. We could also support tensorflow and onnx through nntrainer, but would we want to support other formats too?

@ksh8281 I would really appreciate your feedback!