[webnn] Add float32 tests for resample2d (downsample)

[BruceDai]

This PR is split from #41811, let's keep this one for further investigation and discussion.

Move last discussion on #41811 here.

@fdwr This PR is to add 13 float32 resample2d tests. I ran these tests by internal Chromium build using WebNN DirectML backend, got 10 PASS and 3 FAIL.

These 3 FAIL tests are all for resample(down sampling) on "nearest-neighbor" mode with below 4D input of shape [1, 1, 4, 6]

// nchw layout
[[[[ 3.86005284, 45.18463077, 87.67153743, 98.78210347, 66.37414347,  3.41158349],
   [86.14930501, 95.9813372 , 76.87126314, 16.52591355, 65.98782867, 25.47092156],
   [22.56010548, 92.08479613, 85.80876635, 92.63166027, 29.91620871, 75.40461275],
   [62.06375451,  1.77121588, 99.47231285, 11.44055014, 25.39634271, 67.02175102]]]]

by options.sizes = [2, 3] or options.scales = [0.5, 0.5],

output is of shape [1, 1, 2, 3] and its actual value as below

// using bottom right data
[[[[95.9813372 , 16.52591355, 25.47092156],
   [ 1.77121588, 11.44055014, 67.02175102]]]]

while our expected data were generated by https://github.com/webmachinelearning/webnn-baseline/blob/main/src/resample2d.js with same input data and options:

// using top left data
[[[[ 3.86005284, 87.67153743, 66.37414347],
   [22.56010548, 85.80876635, 29.916208715]]]]

I'm not sure which is wrong on

1. Pure JavaScripts reampled2d implementation for WebNN Baseline
2. Chromium implementation for reampled2d on DirectML backend

Do you have any ideas? Thanks.

And @fdwr commented

I'm not sure which is wrong on

Both are reasonable and not really "wrong" - we just have to pick one. The only truly wrong ones are those libraries that always floor pixels (broken libraries) which result in shifted images, but I believe most image resampling API's round input coordinate on halves toward negative infinity, and ONNX's Resize defaults nearest_mode to round_prefer_floor.

This is easy to control via DML_RESAMPLE2_OPERATOR_DESC in Feature Level 5.1 with the RoundingDirection (used here in the ORT DML EP), but I'm concerned how to achieve it with the older RESAMPLE1 🤔, as my previous comment in the ORT DML EP said that "round_prefer_floor" could not be achieved without an API update. The older function used floor(input.x + 0.5), but that's effectively rounding halves up! ⏳...

[fdwr]

👍