Enable correct RoiAlign max mode with env var

[tianleiwu]

Description

Re-implements the changes from PR #7354 with additional improvements. Fixes #6921 and #6146.

Problem

The RoiAlign operator's original max mode implementation used max(w1d1, w2d2, w3d3, w4d4) — taking the max of individually weighted pixel values per sample point. While this theoretically matches the ONNX reference implementation, an alternative approach — bilinear interpolation first (w1d1 + w2d2 + w3d3 + w4d4), then max across sample points — is arguably more correct for applications.

Additionally:

• The ROI size clamping (max(roi_width, 1)) for non-half_pixel mode was present in both CPU and CUDA providers and incorrectly rejected zero-size ROIs that should produce valid point sampling.
• roi_bin_grid_h/w values were not clamped to a minimum of 1, which could cause issues with certain ROI configurations.

Solution

• Default behavior: Keeps the ONNX spec mathematical behavior — max(w1d1, w2d2, w3d3, w4d4) for max mode pooling. This matches the existing CUDA behavior and ensures ONNX conformance tests pass on CPU.
• Environment variable ORT_ROIALIGN_MAX_USE_BILINEAR_INTERPOLATION=1: Enables bilinear interpolation max mode on the CPU Execution Provider, matching PyTorch/Detectron2 behavior.
• Removes the max(roi_width, 1) / max(roi_height, 1) clamping that prevented zero-size ROI point sampling (fixed in both CPU and CUDA providers).
• Adds roi_bin_grid_h/w clamping to a minimum of 1 in both CPU and CUDA providers.
• Removes the outdated warning about incorrect max mode summation.
• Skips the mask_rcnn_R_50_FPN_1x model tests (model was generated with an incorrect RoiAlign implementation).

Note on CUDA Execution Provider:
The CUDA provider has been fully updated in this PR to align precisely with the CPU provider. It now supports the ORT_ROIALIGN_MAX_USE_BILINEAR_INTERPOLATION environment variable toggle. Additionally, opsets 10-15 and opset 16 have been properly registered, resolving silent fallbacks to the CPU Execution Provider for newer generic models.

Changes

File	Change
roialign.h	Added use_max_bilinear_interp_ member, env var parsing via ParseEnvironmentVariableWithDefault, removed outdated warning
roialign.cc (CPU)	Removed ROI size clamping, added roi_bin_grid min-1 clamping, added conditional max mode (ONNX spec vs bilinear interp)
roialign_impl.cu (CUDA)	Removed ROI size clamping, added roi_bin_grid min-1 clamping, and fully implemented the ORT_ROIALIGN_MAX_USE_BILINEAR_INTERPOLATION toggle down to the interpolation device function.
cuda_execution_provider.cc	Added opsets 10-15 (versioned) and opset 16 kernel registrations for RoiAlign.
roialign_test.cc	Updated MaxModePositive expected values for ONNX spec, added test_roialign_mode_max (from ONNX test suite), added scoped env var tests for both modes, all operational on CPU and CUDA.
TestCase.cc	Skip mask_rcnn_R_50_FPN_1x
skip_model_tests.h	Skip mask_rcnn_R_50_FPN_1x_opset10 on WinML

Test Results

All 13 RoiAlign tests pass:

[  PASSED  ] 13 tests.
  RoiAlignTest.AvgModePositive
  RoiAlignTest.AvgModePositive_half_pixel
  RoiAlignTest.AvgModePositive_output_half_pixel
  RoiAlignTest.OnnxTest
  RoiAlignTest.MaxModePositive
  RoiAlignTest.MaxModePositive_test_roialign_mode_max
  RoiAlignTest.AvgModeNegativeInvalidMode
  RoiAlignTest.AvgModeNegativeSamplingRatio
  RoiAlignTest.AvgModeNegativeInvalidNumRoiDims
  RoiAlignTest.AvgModeNegativeInvalidSecondRoiDims
  RoiAlignTest.MismatchNumRois
  RoiAlignTest.MaxModePositive_test_roialign_mode_max_bilinear_interp
  RoiAlignTest.MaxModePositive_test_roialign_mode_max_env_off

[github-actions]

You can commit the suggested changes from lintrunner.

[github-actions]

Suggested change

	#define ADD_VERSIONED_TYPED_ROIALIGN_OP(T) \
	ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
	RoiAlign, \
	kOnnxDomain, \
	10, \
	15, \
	T, \
	kCudaExecutionProvider, \
	(*KernelDefBuilder::Create()) \
	#define ADD_VERSIONED_TYPED_ROIALIGN_OP(T) \
	ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
	RoiAlign, \
	kOnnxDomain, \
	10, \
	15, \
	T, \
	kCudaExecutionProvider, \
	(*KernelDefBuilder::Create()) \

[github-actions]

Suggested change

	#define ADD_TYPED_ROIALIGN_OP(T) \
	ONNX_OPERATOR_TYPED_KERNEL_EX( \
	RoiAlign, \
	kOnnxDomain, \
	16, \
	T, \
	kCudaExecutionProvider, \
	(*KernelDefBuilder::Create()) \
	#define ADD_TYPED_ROIALIGN_OP(T) \
	ONNX_OPERATOR_TYPED_KERNEL_EX( \
	RoiAlign, \
	kOnnxDomain, \
	16, \
	T, \
	kCudaExecutionProvider, \
	(*KernelDefBuilder::Create()) \

[github-actions]

Suggested change

	#define SPECIALIZED_COMPUTE(T) \
	REGISTER_KERNEL_TYPED(T) \
	ADD_VERSIONED_TYPED_ROIALIGN_OP(T) \
	ADD_TYPED_ROIALIGN_OP(T) \
	#define SPECIALIZED_COMPUTE(T) \
	ADD_VERSIONED_TYPED_ROIALIGN_OP(T) \
	ADD_TYPED_ROIALIGN_OP(T) \

[github-actions]

Suggested change

	const bool half_pixel, \
	const bool half_pixel, \