Support `ihwo` filterLayout for `conv2d` op

[Honry]

In Mobilenet V2 (TFlite), its DepthwiseConv2D op's weights is of ihwo filterLayout, which is not supported in current WebNN spec.

Look at following example, the input of DepthwiseConv2D node is 1x112x112x32 with 'nhwc' layout, its weights is 1x3x3x32, after relu we got its output 1x112x112x32. That means the weights' layout should be 'ihwo'.

[BruceDai]

Hi @Honry, when I converted DEPTHWISE_CONV2D CTS tests of Android NNAPI, I transposed filter data and reshaped filter dimension.

Android NNAPI DEPTHWISE_CONV2D op describes likes following:

0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
1: A 4-D tensor, of shape [1, filter_height, filter_width, depth_out], specifying the filter. 
...
11: An optional ANEURALNETWORKS_BOOL scalar, default to false. 
     Set to true to specify NCHW data layout for input0 and output0. Available since NNAPI feature level 3.

referring to WebNN API conv2d Spec.

A depthwise conv2d operation is a variant of grouped convolution, used in models like the MobileNet, 
where the options.groups = input_channels = output_channels and the shape of filter tensor is
[options.groups, 1, height, width] for "nchw" layout or [height, width, 1, options.groups] for "nhwc" layout.

to convert Android NNAPI DEPTHWISE_CONV2D tests to WebNN CTS tests, if the layout of input0 is default 'NHWC', I converted filter with [filter_height, filter_width, 1, depth_out] ('hwio') dimension, else the layout of input0 is 'NCHW', I converted filter with [depth_out, 1, filter_height, filter_width] ('oihw') dimension.

[huningxin]

@Honry and @BruceDai, thanks for the info. It looks like WebNN needs to support ihwo filter layout of conv2d. Then it would not need to convert the filter layout when run Android NNAPI DEPTHWISE_CONV2D test case on WebNN.