pjrt/oneapi: Add support for oneAPI in GitHub workflows and related scripts

[neudinger]

in https://github.com/zml/xla/tree/zml/oneapi

• Basic oneAPI/SYCL single-device support:

• Wires SYCL BLAS through oneMKL.

• Adds GEMM and strided-batched GEMM support.

• Supports several dtypes: FP16, BF16, FP32, FP64, complex, and some int8 paths.

• Adds basic BLAS-Lt matmul execution, including default and bias epilogues.

• Makes oneCCL multi-device collectives explicitly unsupported instead of returning fake success.

• Fixes SyclStream::RecordEvent so it actually updates the passed SyclEvent.

• Fixes oneAPI module caching:

• Replaces raw SPIR-V buffer pointer cache keys with stable SPIR-V byte fingerprints.

• Identical SPIR-V binaries loaded from different host buffers now share the same cached Level Zero module.

• Adds refcount/cache hit/cache miss logging.

• Adds timing logs for Level Zero module creation.

• Adds a test that loads copied SPIR-V bytes and verifies both loads return the same module handle.

• oneAPI/SYCL BLAS functional for single-device execution.

• SPIR-V module loading avoids duplicate Level Zero module creation for identical binaries.

CUDA and ROCm already cache loaded modules behind ModuleHandle/refcount maps,
but key them by the raw binary pointer. For oneAPI, i use a stable fingerprint
of the SPIR-V bytes instead, so identical SPIR-V buffers at different host addresses share the same loaded Level Zero module.