[CUDA] Add build option for TensorRT fused attention cubins

[Tianlei Wu (tianleiwu)]

This change adds a build option for excluding the prebuilt TensorRT fused multi-head attention cubins from the CUDA Execution Provider. The option is enabled by default to preserve existing behavior; disabling it removes approximately 14 MB of embedded cubin data and leaves attention selection to the other available kernels and unfused fallback paths.

Key Changes

• Add onnxruntime_USE_TRT_FUSED_ATTENTION, dependent on CUDA and enabled by default.
• Propagate the option to CUDA provider compilation and guard TensorRT fused-attention cubin declarations, metadata, and lookup paths.
• Filter the prebuilt TensorRT fused-attention cubin sources from both the regular CUDA provider and CUDA plugin provider when the option is disabled, while retaining the shared driver wrapper needed by sparse attention.
• Add Windows plugin DLL version metadata and suppress build-machine RPATH embedding for packaged Linux plugin binaries.
• Exercise onnxruntime_USE_TRT_FUSED_ATTENTION=OFF in the Windows CUDA no-cuDNN plugin build.

Testing Notes

• git diff --check origin/main...HEAD passes.
• The Windows CUDA no-cuDNN workflow now builds the CUDA plugin with TensorRT fused attention disabled, providing CI coverage for the opt-out configuration.
• No local CUDA/Windows build was run in this environment.

[Copilot]

Pull request overview

Adds a CUDA build-time opt-out for embedding TensorRT fused multi-head attention prebuilt cubins, reducing binary size when disabled while preserving current defaults and ensuring the plugin build remains package-friendly across platforms.

Changes:

• Introduce onnxruntime_USE_TRT_FUSED_ATTENTION (CUDA-dependent, default ON) and propagate it as USE_TRT_FUSED_ATTENTION to the provider build.
• Guard TensorRT fused-attention cubin declarations/metadata/lookup helpers and filter the corresponding .cubin.cc / _kernel.sm*.cc sources when the option is OFF.
• Update CUDA plugin packaging behavior (Windows DLL version resource; Linux skip build RPATH) and add CI coverage for the opt-out configuration.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/mha_runner.cu	Stubs TRT fused runner support when the build option is disabled.
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/fused_multihead_attention_v2.h	Guards TRT fused cubin metadata/lookup and provides null-return stubs when disabled.
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/flash_attention/fmha_flash_attention.h	Guards flash-attention cubin metadata/lookup behind the build option.
onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/cross_attention/fmha_cross_attention.h	Guards cross-attention cubin metadata/lookup behind the build option.
cmake/onnxruntime_providers_cuda.cmake	Filters out TRT fused cubin blob sources when the option is OFF.
cmake/onnxruntime_providers_cuda_plugin.cmake	Applies the same filtering for plugin builds; adds Windows version resource and Linux RPATH suppression.
cmake/CMakeLists.txt	Defines the new onnxruntime_USE_TRT_FUSED_ATTENTION option and emits the provider compile define when enabled.
.github/workflows/windows_cuda_no_cudnn.yml	Exercises onnxruntime_USE_TRT_FUSED_ATTENTION=OFF in Windows CUDA plugin CI.