Add checks for model spec and matching output values in `to_onnx()` method

🚀 Feature

The official PyTorch tutorial for exporting to ONNX includes checking the model spec as well as confirming the output values match the PyTorch version of the model. The current Lightning implementation of to_onnx() only calls torch.onnx.export but skips these additional checks. For best practice, I think they should be included (if not by default, then with an optional boolean flag to turn them on)

Motivation

1. Follow the official tutorial as closely as possible
2. Ensure the ONNX version of the model does not produce unexpected results before it's used in production

Pitch

Add something along the lines of the following code to the end of the current to_onnx() method

import onnx
import onnxruntime
import numpy as np

def to_numpy(tensor):
    return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()

onnx_model = onnx.load(file_path)
onnx.checker.check_model(onnx_model)
ort_session = onnxruntime.InferenceSession(file_path)
ort_inputs = {ort_session.get_inputs()[0].name: to_numpy(input_sample)}
ort_outs = ort_session.run(None, ort_inputs)
for ort_out, torch_out in zip(kwargs["example_outputs"], ort_outs):
    np.testing.assert_allclose(to_numpy(torch_out), ort_out, rtol=1e-03, atol=1e-05)

Alternatives

Users override the method and implement the checks themselves. However, I think these are general enough that they should be included in the base method to limit boilerplate code

Additional context

• This would require including the Python bindings for onnx and onnxruntime with the Lightning distribution
• The biggest challenge to generalize would be the model outputs. Many models have more than one output node, so we need to make sure each node is compared with the appropriate corresponding ONNX node. Iterating with zip is probably alright unless the PyTorch version outputs a dict (which I think would be bad practice)
• In my own use, I have found you need to be very careful about the order of library imports. Details in this issue comment from the onnx repo