Provide a higher-level public API for Rust users

[abrown]

The functions defined in wasi_nn::wasi_ephemeral_nn::* are the raw Wasm symbols used to link with the host's implementation of wasi-nn. The wasi-nn Rust bindings should not expose these, since they should be providing a higher level of abstraction. These symbols are visible in the documentation here.

[abrown]

The following is a sketch of what the higher-level API could look like:

struct Graph { ... }
impl Graph {
  pub fn load(data: impl Iterator<Item = &[u8]>, encoding: GraphEncoding, target: ExecutionTarget) -> Result<Self>;
  pub fn get_execution_context(&self) -> Result<ExecutionContext>;
  pub fn get_input_types(&self) -> TensorTypes;
  pub fn get_output_types(&self) -> TensorTypes
}

struct TensorTypes { ... }
impl TensorTypes {
  pub fn len(&self) -> usize;
  pub fn get(&self, index: u32) -> Result<TensorType>;
}

struct ExecutionContext { ... }
impl ExecutionContext {
  pub fn set_input(&mut self, index: u32, tensor: Tensor) -> Result<()>;
  pub fn compute(&mut self) -> Result<()>;
  pub fn get_output(&self, index: u32, index: u32) -> Result<Tensor>;
}

Some considerations:

• perhaps get_execution_context must be a &mut self, but it would be preferable if it were not
• ExecutionContext may need a lifetime parameter since a Tensor may have an associated lifetime
• we should not make get_output consume self since there may be multiple outputs that need to be retrieved
• note how none of the APIs need be unsafe
• to support introspecting the graph inputs and outputs, the wasi-nn spec will need some additions (e.g., get_input_types_len, get_input_type, get_output_types_len, get_output_type)

[abrown]

The functions defined in wasi_nn::wasi_ephemeral_nn::* are the raw Wasm symbols used to link with the host's implementation of wasi-nn. The wasi-nn Rust bindings should not expose these, since they should be providing a higher level of abstraction. These symbols are visible in the documentation here.

Once new bindings are created, the old low-level symbols should be removed from the public API.

[geekbeast]

We should also address https://github.com/bytecodealliance/wasi-nn/issues/13 as well. Being able to generate a compatible tensor server side seems like it will be the entry point to most workflows.

My initial thoughts are that wasi-nn shouldn't care about correctly implementing the mapping from images onto tensors, but we may want to provide a conversion api for bf16 if it is commonly used. I see no reason we can't provide helper functions for extracting tensor from the image, but they shouldn't be part of the spec.

[abrown]

We should also address https://github.com/bytecodealliance/wasi-nn/issues/13 as well.

@geekbeast, commented over in that issue.