[ET-VK] native_layer_norm rejects a missing weight, so affine-free F.layer_norm cannot run

[msluszniak]

🐛 Describe the bug

add_native_layer_norm_node throws when weight is None:

if (graph.val_is_none(weight_data)) {
  VK_THROW("native_layer_norm requires weight to be non-None");
}

The neighbouring code already handles a missing bias, synthesizing a zero-bias TensorRef for nn.LayerNorm(bias=False). A missing weight is the symmetric case and is just as ordinary: F.layer_norm(x, (channels,), eps=...) with no affine parameters is a normal call, used whenever a module applies its own scale and shift afterwards.

Concrete case: kokoro's AdaLayerNorm normalizes with no affine parameters and then applies a style-conditioned (1 + gamma) * x + beta. Lowering that model to Vulkan produces a .pte that aborts at prepack:

Exception raised from add_native_layer_norm_node at
backends/vulkan/runtime/graph/ops/impl/NativeLayerNorm.cpp:95:
native_layer_norm requires weight to be non-None

The partitioner claims the op, so there is no way to avoid this from the export side other than blocklisting native_layer_norm entirely, which splits the graph.

A unit weight and a zero bias reproduce out = (x - mean) * rstd exactly, so synthesizing the missing parameter is enough. When there is no weight tensor to take a shape from, normalized_shape supplies it.

Versions

Reproduces on main (a6b115b).