Is there a reason to support out of order axes for layerNormalization?

[philloooo]

Hi!
I found that on CoreML, when the axes are out of order, sometimes the result is wrong. I've also tested with tensorflow, which also returns wrong result when axes are out of order

Is there a reason to support unordered axes?

We could emulate this by transposing(it's just not as efficient, CoreML only support scale and bias as const , so applying transposed bias and scale have to be done outside of layernorm), but if there is no valid use case to support unordered axes, it's simpler to just reject that.