sycl: add REPEAT_BACK operation support

[shani-f]

Summary

Implements the REPEAT_BACK operator for the SYCL backend, enabling mirrored repetition of tensor data on SYCL devices.
The implementation matches the behavior of the existing CPU version.

Changes

• Added repeat_back.cpp with SYCL kernel implementation
• Added dispatch in ggml-sycl.cpp under GGML_OP_REPEAT_BACK
• Integrated into the SYCL compute pipeline

Implementation

• Supports tensors up to 4 dimensions
• Performs reflection-based repetition using modular indexing
• Parallelized with parallel_for using range<3>
• Efficient indexing for mirrored boundaries
• Zero-overhead path for non-repeated cases

Testing

• Verified results match the CPU implementation
• Checked correctness on multi-dimensional tensors
• Tested reflection accuracy on all tensor edges

Performance

• Optimized index computation (no redundant divisions)
• Memory accesses follow contiguous layout
• Minimal kernel footprint and register usage

Compatibility

• Supports F32 tensors
• Works with Level Zero and OpenCL backends
• Follows existing SYCL backend design conventions

[shani-f]

@CISC @NeoZhangJianyu
this PR adds SYCL backend support for GGML_OP_REPEAT_BACK.
All tests pass locally and the implementation currently supports F32 tensors.

[shani-f]

Hello @CISC @NeoZhangJianyu,
I noticed that the two failing checks are WebGPU and WebUI tests, which seem unrelated to my changes.
I’d appreciate it if you could let me know what the next step is toward merging.
Thank you!

[CISC]

I noticed that the two failing checks are WebGPU and WebUI tests, which seem unrelated to my changes. I’d appreciate it if you could let me know what the next step is toward merging. Thank you!

It happens sometimes, I've re-run them, but don't worry about it. @NeoZhangJianyu will review when time permits.