[XLA:GPU][oneAPI] Preview oneAPI 2026.0 toolchain

[neudinger]

📝 Summary of Changes

This draft PR previews the XLA-side integration for oneAPI 2026.0 hermetic builds.

It temporarily pins rules_ml_toolchain to the head of:

• [XLA:GPU][oneAPI] Add oneAPI BUILD template and update redist versions for 2026.0 google-ml-infra/rules_ml_toolchain#271

Once that PR lands, this PR will be updated to use the official google-ml-infra/rules_ml_toolchain commit instead of the temporary fork archive.

🎯 Justification

Same as google-ml-infra/rules_ml_toolchain#271 :

• To update ZML to fully support the latest intel GPU and the lattest oneAPI toolkit, we need to update the XLA project with oneAPI 2026.

Future oneAPI development in XLA needs access to the latest Intel toolkit and library layout :

• Enables the hermetic SYCL configuration to select oneAPI 2026.0 on Ubuntu 24.04 and updates the reported PJRT platform version accordingly.
• This PR intentionally does not include runtime changes, SYCL BLAS work, TopK changes, PJRT synchronization changes, oneCCL changes, or packaging changes.

So that we can land the oneAPI 2026.0 update and unblock future development that depends on it, while minimizing the risk of breaking changes in those other areas.

Note :

The temporary rules_ml_toolchain archive depends on the redist source described in google-ml-infra/rules_ml_toolchain#271.
As explained in that PR, the redist source is not published by Intel.

Intel does not currently provide standalone redistributable packages for this oneAPI release. They only provide the toolkit installer:
https://www.intel.com/content/www/us/en/developer/tools/oneapi/oneapi-toolkit-download.html?packages=oneapi-toolkit&oneapi-toolkit-os=linux&oneapi-lin=offline
For now, this PR uses redistributable packages hosted here:
https://github.com/neudinger/rules-ml-toolchain-redists

I prefer to replace this package with an Intel-hosted alternative when one becomes available.

cc @bhavani-subramanian @akhilgoe @nhatleSummer22 @ashiqimranintel

🚀 Kind of Contribution

📊 Benchmark (for Performance Improvements)

🧪 Unit Tests:

🧪 Execution Tests:

bazel build //xla/pjrt/c:pjrt_c_api_gpu_plugin.so --config=sycl_hermetic --nobuild
bazel build //xla/stream_executor/sycl/... --config=sycl_hermetic --nobuild
bazel build //xla/pjrt/c:pjrt_c_api_gpu_plugin.so --config=sycl_hermetic