Switch the exporter to aten IR

[gedoensmax]

The FX graph parsing was a bit too much. So i lowered to aten which seems to be a good compromise.
I also added some updates to the flux sample.

[gedoensmax]

@mtavenrath I think the unit tests failures are not a failure in the exporter.

For Conv i am getting this error:

tests/conftest.py:151: in validate_webnn_execution
    if not torch.allclose(exp, actual, rtol=rtol, atol=atol):
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E   RuntimeError: The size of tensor a (26) must match the size of tensor b (24) at non-singleton dimension 3

This indicates that there is incorrect padding parsing or forwarding to ONNX in the graph deserialization of below graph:

webnn_graph "model" v1 {
  inputs { x: f32[1, 1, 28, 28]; }
  consts {
	weight_1: f32[32, 1, 5, 5] @weights("conv.weight");
	weight_2: f32[32] @weights("conv.bias");
  }
  nodes {
	[operand_4] = reshape(weight_2, newShape=[1, 32, 1, 1]);
	[operand_5] = conv2d(x, weight_1, filterLayout="oihw", groups=1, inputLayout="nchw", pads=[1, 1, 1, 1]);
	[operand_3] = add(operand_5, operand_4);
  }
  outputs { operand_3; }
}

[gedoensmax]

The concat operators seems to have issues as well:

webnn_graph "model" v1 {
  inputs { tensors_0: f32[2, 3, 4, 5]; tensors_1: f32[2, 3, 4, 5]; }
  nodes {
	[operand_1] = concat([tensors_0, tensors_1], axis=0);
  }
  outputs { operand_1; }
}

Runtime error:

E   RuntimeError: ONNX execution failed: onnx runtime failed: load model failed: This is an invalid model. In Node, ("concat_0", Concat, "", -1) : () -> ("operand_1": tensor(float),) , Error Node(concat_0) with schema(::Concat:13) has input size 0 not in range [min=1, max=2147483647].

[mtavenrath]

There is a PR which fixes several webnn-graph related issues here rustnn/webnn-graph#10. Can you give it a try?