Upgrade Struct2Tensor to use TensorFlow 2.18.1

[vkarampudi]

Pull Request: Upgrade to TensorFlow 2.18.1 & Fix CI Builds

This PR upgrades struct2tensor to TensorFlow 2.18.1 and resolves persistent CI build failures on both Ubuntu and macOS runners. It removes environment-specific hardcoded paths and introduces robust workarounds for build system incompatibilities.

🚀 Key Changes & Justifications

1. Workspace & Dependency Upgrades

WORKSPACE
Change: Upgraded _TENSORFLOW_GIT_COMMIT to point to TF 2.18.1 and filled in _TENSORFLOW_ARCHIVE_SHA256 (5fcca4ec0732e146547fc26f6966cf92236c2f85a9e086b85c83d303e2c38980).
Justification: To release version 0.49.0 supporting TensorFlow 2.18.1.
Change: Replaced
tensorflow.patch
with a new
proto_gen.patch
.
Justification: The old patch was specific to TF 2.17.1. The new patch is required to fix protocol buffer generation path mapping bugs in the new TF internal structure.
Change: Added local_repository for local_config_cuda pointing to third_party/cuda_stub.
Justification: TensorFlow's workspace requires this definition even if building for CPU-only. It satisfies the dependency graph without installing full CUDA.

environment.yml
Change: Upgraded pyarrow from 10.0.1 to 15.0.0.
Justification: Resolved installation failures on Apple Silicon (arm64) macOS runners.
Change: Fixed tensorboard version from 2.18.1 (which does not exist) to 2.18.0.
Justification: Resolved environment creation failures on CI.

2. CI Environment Independence

.bazelrc
Change: Removed hardcoded local python/tensorflow paths (e.g., pointing to local Anaconda environments of the developer).
Justification: Hardcoded paths caused the CI to search for directories that only existed on the developer's machine. The ./configure.sh script is now the standard mechanism to generate these paths dynamically on the runner.

3. File Cleanup & Additions

📂 third_party/cuda_stub
Change: [NEW] Added a lightweight stub repository.
Justification: Acts as a placeholder to allow building without CUDA dependencies. Required by WORKSPACE to satisfy the TensorFlow graph on CPU-only runners.

third_party/tensorflow.patch
Change: [DELETE] Removed the old patch.
Justification: Dead code, superceded by the new proto_gen.patch.
✅ Verification
The build process succeeds without hardcoded errors. You can verify by running the tests inside your conda environment.

conda run -n s2t-env ./configure.sh conda run -n s2t-env bazel build //...