Give Outputs Direct Access to Tensor Data

[Michael-F-Bryan]

Copied straight from #242 (comment) - essentially #242, but generalised to all outputs.

---

My thoughts are that when the Rune requests an output it should provide some sort of specifier/descriptor which contains things like what type of data the output should expect and any extra arguments (e.g. a human-friendly name to help differentiate outputs or a wallet hash).

{
  "destination": "SERIAL",
  "types": ["str[1024]", "f32[1024]"],
  "args": {
    "description": "Labels and their confidence values"
  }
}

(this only gets run once on startup so I think it's okay to deserialize JSON instead of using something more strongly typed/complex)

Then consuming output would be made to look a bit like rune_model_infer() where you provide a pointer to an array of pointers to tensor data in WebAssembly linear memory. The output combines that with the type information provided while loading to figure out what tensors it's been given and how to deal with them.

Our SERIAL output will unconditionally serialize the tensor to JSON and gives the runtime a reference to the resulting string, so while it may be more convenient for the mobile and web applications, it's also quite limiting and the serializing process incurs a substantial performance penalty.