Dynamic dimension support, fix multiple inputs issue

[nbigaouette]

1. Add support for dynamic dimension

   Dynamic dimension is supported in ONNX and is stored as -1 in the model file. Use an Option<> for the dimension values so None can represent "unknown" (dynamic). Dimension of the input array is used to fix the size to send to the runtime.

2. Fix problem with multiple inputs (and outputs)

   Don't loop over the input values and call Run() multiple times. Multiple inputs means the model takes multiple inputs to perform a single inference. Pack all inputs together before sending to a single Run() call.

3. Fix bug with input and output types

   Initially inputs and outputs generic types were the same. This is not necessarily true. Inputs and outputs should be able to have different types. A new generic type is added to Session::run() for the output type.

Closes #22.

[bminixhofer]

LGTM! I tried the following, adapted from the issue:

Cargo.toml

[package]
name = "onnxruntimerstest"
version = "0.1.0"
authors = ["Benjamin Minixhofer <bminixhofer@gmail.com>"]
edition = "2018"

# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html

[dependencies]
onnxruntime = { git = "https://github.com/nbigaouette/onnxruntime-rs", branch = "issue22" }
ndarray = "0.13.1"

src/main.rs

use ndarray::Array2;
use onnxruntime::environment::Environment;
use onnxruntime::tensor::OrtOwnedTensor;
use onnxruntime::GraphOptimizationLevel;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let env = Environment::builder().with_name("env").build()?;
    let mut session = env
        .new_session_builder()?
        .with_optimization_level(GraphOptimizationLevel::Basic)?
        .with_model_from_file("model.onnx")?;

    println!("{:#?}", session.inputs);
    println!("{:#?}", session.outputs);

    let input_ids = Array2::<i64>::from_shape_vec((1, 3), vec![1i64, 2i64, 3i64])?;
    let attention_mask = Array2::<i64>::from_shape_vec((1, 3), vec![1i64, 1i64, 1i64])?;

    let output: OrtOwnedTensor<f32, _> = session
        .run(vec![input_ids.clone(), attention_mask])?
        .remove(0);

    println!("{:#?}", output.into_slice().unwrap());

    Ok(())
}

I get the same output as with equivalent onnxruntime Python code and the input / output dims and datatypes look correct to me.