litert/xnnpack: add rank guards to AveragePool2D, MaxPool2D, Conv2D, DepthwiseConv2D, and TransposeConv2D ToXnnpack() before shape subscripts

[destro4evr-rgb]

Five ToXnnpack() implementations access fixed shape indices (shape[1], shape[2], shape[3]) on input, filter, and output tensor shape vectors without first validating that the vector has enough elements. A crafted TFLite model that declares a wrong-rank tensor for any of these operations triggers a null pointer dereference during XNNPACK subgraph compilation at model load time, before any inference runs.

TransposeConvOperation::ToXnnpack() (lines 1649–1653) performs the same shape subscripts and already includes the explicit rank guard that the five affected functions are missing:

if (input_info.shape.size() != 4 || filter_info.shape.size() != 4 ||
    output_info.shape.size() != 4) {
    return absl::InvalidArgumentError(...);
}

Fix: add equivalent rank size checks before the first shape subscript in each affected function — size() < 3 for AveragePool2D and MaxPool2D (which only need indices 1 and 2), size() != 4 for Conv2D and DepthwiseConv2D (input + filter), and size() != 4 for all three tensors in TransposeConv2D.