Support all tensor types

[radu-matei]

Currently, if input and output tensors are not F32, both runtimes will panic.
This stems from the way we handle and store input and output tensors, particularly when using the native ONNX runtime - where both input and output tensors are Vec<Array<f32, Dim<IxDynImpl>>>.

The additional tensor types defined by WASI NN are F16, U8, and I32.

We need to either:

• make both tensors generic over the data type. This would potentially make the entire State struct generic, which could introduce difficulties in how we manage it.
• keep storing every tensor type as F32, introduce fields for input and output tensor types (Question: how to handle input model not containing tensor shapes? #12) and transform them just-in-time (when compute is called, for input tensors, and when get_output is called, for output tensors). This would however introduce an additional performance penalty at inference time.

[Mossaka]

When I get back from vacation I will look into this.