GlobalAveragePool infers the input's spatial dims into the output static_shape

[antimora]

Summary

GlobalAveragePoolProcessor::infer_types sets the output type by cloning the input's TensorType, which copies the input's spatial dimensions into the output's static_shape. Per the ONNX spec every spatial dim of the output is 1.

The comment on the line already states the intended behavior:

// Output has the same type and rank as input (spatial dimensions become 1)
node.outputs[0].ty = ArgType::Tensor(input_tensor.clone());

crates/onnx-ir/src/node/global_avg_pool.rs:56-57

Reproduction

A GlobalAveragePool on a [1, 3, 8, 8] input, converted with onnx2burn. From the generated debug dump:

name: "globalaveragepool1_out1",
ty: Tensor(TensorType {
    dtype: F32,
    rank: 4,
    static_shape: Some([Some(1), Some(3), Some(8), Some(8)]),

Expected [Some(1), Some(3), Some(1), Some(1)].

Impact

Rank and dtype are correct, so this does not break the node itself. It misleads anything that trusts static_shape: shape-dependent constant folding, Reshape/Expand/Slice inference, and the partitioner's shape reasoning.

I did not find an actual miscompile in a quick probe. A following Shape node falls back to a runtime dims() read rather than folding the static shape, so it does not surface there. Filing as a latent correctness issue rather than an active one.

Suggested fix

The same rule GlobalLpPool implements in crates/onnx-ir/src/node/global_lp_pool.rs: keep N and C, set every dim from index 2 on to Some(1), and take the length from rank so the shape cannot disagree with the rank field beside it. With two call sites, a small shared helper may be worth extracting.