[ONNX] Fix test_roi_align failure

[sfvaroglu]

RoiAlign-16 introduces coordinate_transformation_mode, which should be set to 'output_half_pixel' to omit the pixel shift for the input (for a backward-compatible behavior). This PR should fix the failure in https://ci.tlcpack.ai/job/docker-images-ci/job/docker-image-run-tests/231/testReport/junit/cython.tests.python.frontend.onnx/test_forward/Test___frontend__GPU_3_of_6___test_roi_align_cuda_/

[sfvaroglu]

@leandron @driazati

[sfvaroglu]

The roi_align test passes locally when using onnx==1.12.0 and onnxruntime==1.12.1, but it looks like CI is failing with onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : This is an invalid model. Error in Node: : Unrecognized attribute: coordinate_transformation_mode for operator RoiAlign which means the CI image doesn't have the updated onnx and onnxruntime, but the CI image at https://ci.tlcpack.ai/job/docker-images-ci/job/docker-image-run-tests/231/testReport/junit/cython.tests.python.frontend.onnx/test_forward/Test___frontend__GPU_3_of_6___test_roi_align_cuda_/ has the updated onnx and onnxruntime and fails with a different error.

What do we need to update the CI image this PR is running on?

[driazati]

Since we need to update the image to fix the test and fix the test to update the image, https://github.com/apache/tvm/tree/sevin/test_roi_align_fix rolls them all into one (CI jobs that run against branches in apache/tvm will use the newly specified images). #12906 passed the branch CI and hopefully will pass the other so we can merge that one