[Dup] Fix SAME_UPPER/SAME_LOWER (auto_pad attribute) in ConvTranspose

[jcwchen]

Description:
Dup of #5368. Recreate this PR for cleaner commit log. Target it to ORT 1.13. Recreating this PR to retarget the main branch instead of master branch.

To sync the definition of SAME_UPPER/SAME_LOWER among all operators and make it same as ONNX definition, switch the logic of SAME_UPPER and SAME_LOWER in ConvTranspose.

Definition of SAME_UPPER and SAME_LOWER should be as follows:

if auto_pads == 'SAME_UPPER':
  pad_head = paddings / 2  # smaller one 
  pad_tail = paddings - paddings / 2  # larger one
elif auto_pads == 'SAME_LOWER':
  pad_head = paddings - paddings / 2  # larger one
  pad_tail = paddings / 2  # smaller one

Besides, revert the temporary warning from" #11984.

Motivation and Context
The auto_pad attribute, SAME_UPPER and SAME_LOWER of ConvTranspose is different from other operators' (pool and conv related operators) auto_pad attribute. The behavior of same attribute should be the same among all operators. Also, it does not meet the definition in ONNX.

• SAME_UPPER and SAME_LOWER in other operators
  

  onnxruntime/onnxruntime/core/providers/cpu/nn/pool_attributes.h

  Line 149 in c20fcf2

  *pad_head = pad_needed / 2;

• ConvTranspose definition in ONNX:
  https://github.com/onnx/onnx/blob/b2ed660d0a065b8346816f2c3a95d79ca79b88c9/onnx/defs/nn/defs.cc#L1222
• A related PR in ONNX (There is a contradiction of this in ONNX):
  Update spec for Convtranspose to make it sync onnx/onnx#3019

[fdwr]

We should confirm that the CPU EP is the only EP affected by this, not also DML, CUDA... Generally seems fine to me, if it was a mistake in implementation per the spec, but I wouldn't want to just fix the CPU EP and now have discrepancies between EP's.

Looking here at ConvolutionHelperBase::InitializeKernelAndShapesTransposed, the DML EP may have the same issue.

[hariharans29]

LGTM. Please run this by @pranavsharma to get a sign-off as well.

[jcwchen]

We should confirm that the CPU EP is the only EP affected by this, not also DML, CUDA... Generally seems fine to me, if it was a mistake in implementation per the spec, but I wouldn't want to just fix the CPU EP and now have discrepancies between EP's.

Thank @fdwr for the review! I did check CUDA ep and it seems it just uses the CPU implementation to calculate ConvTranspose's attribute see the line here. @hariharans29 Please correct me if I am wrong.

Looking here at ConvolutionHelperBase::InitializeKernelAndShapesTransposed, the DML EP may have the same issue.

Good catch. I think that line is incorrect. By contrast, the line in the same file is right. I can include the fix in this PR.

I roughly went through other EPs as well: ORT web implemented it correctly originally. See this line. Other than that, it seems to me other EPs do not have specific kernel for ConvTranspose's SAME/UPPER/LOWER case.

[fdwr]

FYI @jeffbloo.