Fix RoiAlign CPU EP issues

[fdwr]

Description:
(1) max mode should take the entire bilinear sample, not a fragment of it. This line should be output value = max(sum of all weighted values, previous maximum), not output value = max(max(max(max(value1, value2), value3) value4), previous maximum).
(2) 0 size regions are valid and should be treated as sampled points, not forced to 1. This leads to test failures when sampling tiny points or slivers in our ONNX conformance tests. bin_size_w and bin_size_h are only multiplied anyway, not divided, and so an attempt to prevent division by zero (or whatever the original intention) is unnecessary.
(3) The ONNX compat attribute was already added in onnx/onnx#3625 input pixel alignment is 0.5 off - the correct equation is (x + 0.5) * scale - 0.5, but the final 0.5 was missed (x + 0.5) * scale (see https://arxiv.org/abs/1703.06870).

Motivation and Context

• Why is this change required? What problem does it solve? The DML/WinML ONNX conformance tests fail on CPU for a number of cases.
• If it fixes an open issue, please link to the issue here.
  • Dubious implementation of ROIAlign in 'max' mode #6146
  • related: RoiAlign CPU is not aligned to pixel centers (per the Mask RCNN paper and Facebook's Detectron2 implementation) #6921
  • related: [Operator] RoiAlign backend test case's output is misaligned onnx/onnx#3428.

[jiafatom]

Thanks for referring my PR to this PR. I think we agree that the ultimate solution is to update onnx spec for ROIAlign - we need add an attribute aligned. And based on that, we change onnxruntime logic accordingly -- refer to the PR for aligned=True case. The logic in my PR is in sync with torch ROIAlign implementation (I tested the result matches). So I would prefer to use logic there.

[fdwr]

@hariharans29 - do you have instructions on updating this C# test baseline? I believe these values were recorded with the bug baked into them:

System.Exception : Opset opset10, Model mask_rcnn: ModelFile = D:\a\_work\1\b\models\opset10\mask_rcnn\mask_rcnn_R_50_FPN_1x.onnx error = Assert.Equal() Failure
Expected: DenseTensor<Single> [475.39844, 286.16537, 803.5265, 553.59546, 873.8627, ...]
Actual:   DenseTensor<Single> [475.39844, 286.16534, 803.5265, 553.5954, 873.8627, ...]

https://dev.azure.com/onnxruntime/onnxruntime/_build/results?buildId=674145&view=logs&jobId=ed9f9512-dfe4-5f72-a421-4225d265fdd2&j=ed9f9512-dfe4-5f72-a421-4225d265fdd2&t=d93860a7-6075-51af-35c2-8a3248009377

[hariharans29]

@hariharans29 - do you have instructions on updating this C# test baseline? I believe these values were recorded with the bug baked into them:

System.Exception : Opset opset10, Model mask_rcnn: ModelFile = D:\a\_work\1\b\models\opset10\mask_rcnn\mask_rcnn_R_50_FPN_1x.onnx error = Assert.Equal() Failure
Expected: DenseTensor<Single> [475.39844, 286.16537, 803.5265, 553.59546, 873.8627, ...]
Actual:   DenseTensor<Single> [475.39844, 286.16534, 803.5265, 553.5954, 873.8627, ...]

https://dev.azure.com/onnxruntime/onnxruntime/_build/results?buildId=674145&view=logs&jobId=ed9f9512-dfe4-5f72-a421-4225d265fdd2&j=ed9f9512-dfe4-5f72-a421-4225d265fdd2&t=d93860a7-6075-51af-35c2-8a3248009377

Maybe we could change the test referencing the PR and stating that the expected values were generated with the buggy implementation to begin with ?

[yuslepukhin]

Is there an issue filed with ONNX? We need to reference it here.

[fdwr]

yuslepukhin onnx/onnx#3428

Is there an issue filed with ONNX? We need to reference it here.

Well here was the original ONNX issue onnx/onnx#3428, which RoiAlign opset 16 addressed by adding coordinate_transformation_mode, but did you mean one specific to model generation, for the ONNX model zoo?