[torchlib] Impl `aten::native_group_norm` with `onnx::GroupNormalization`

[BowenBao]

Saw in comment

    # Because onnx.GroupNorm() need size=group for weight and bias
    # But the torch's aten function's input need size=channel, the size mismatched
    # So we have to use onnx.InstanceNorm() to simulate

and previous discussion #644 (comment).

Reviving the thread to see if something can be done to use ONNX GroupNorm. The current workaround with InstanceNorm produces way more nodes which are not ideal for performance.

E.g.:

cc @justinchuby @xiaowuhu

[BowenBao]

Setting to low pri for now since it appears the extra computation is for the extra mean/var outputs.

[justinchuby]

Graph norm is not correctly defined in ONNX. We need to fix it: onnx/onnx#5466

[justinchuby]

We can do this when we start to use Opset 21.

[justinchuby]

Done https://github.com/pytorch/pytorch/blob/733e684b11acaf2b6ff4b4e8c459a9c71e903d81/torch/onnx/_internal/exporter/_torchlib/ops/nn.py#L46