The input of triangular should be at least 2D tensor

[BruceDai]

This is raised by @fdwr on reviewing implementation for triangular op on webmachinelearning/webnn-baseline#64. (Thanks Dwayne!)

it should say the input is at least 2D. You can see other ML libraries just treat each 2D subsection as a separate batch:

https://numpy.org/doc/stable/reference/generated/numpy.triu.html
https://pytorch.org/docs/stable/generated/torch.tril.html
https://onnx.ai/onnx/operators/onnx__Trilu.html

So shape [4,3,5,7] would apply the triangular masking to the last two dimensions.

[fdwr]

...

  Arguments:
- input: an [MLOperand](https://www.w3.org/TR/webnn/#mloperand). The input 2-D tensor.
+ input: an [MLOperand](https://www.w3.org/TR/webnn/#mloperand). The input tensor which is at least 2D.

   options: an optional [MLTriangularOptions](https://www.w3.org/TR/webnn/#dictdef-mltriangularoptions). The optional parameters of the operation.

- Returns: an [MLOperand](https://www.w3.org/TR/webnn/#mloperand). The output 2-D tensor representing a triangular matrix.
+ Returns: an [MLOperand](https://www.w3.org/TR/webnn/#mloperand). The output tensor representing a triangular matrix which is the same shape as the input.

...

+ // lower triangular matrix with two batches:
+ //   [[[7, 0, 0],
+ //     [9, 4, 0],
+ //     [2, 6, 3]],
+ //    [[1, 0, 0],
+ //     [4, 5, 0],
+ //     [7, 8, 9]]]
+ const lower = builder.triangular(input, { upper: false });

[BruceDai]

Thanks @fdwr !
I think you can create a PR for this!