sycl: add RMS_NORM_BACK operation support

[YaelLogic]

Summary

Add SYCL backend support for RMS_NORM_BACK using a single FP32 compensated parallel reduction path.
No changes to the public API. Default numerical accuracy is preserved; a fast opt-in macro is also available.

---

Implementation

Algorithm (onsistent with existing backend behavior)

• inv_r = 1 / sqrt( (Σ x²) / D + eps )
• coeff = − (Σ x·dz) / (Σ x² + D·eps)
• dx[i] = (dz[i] + coeff * x[i]) * inv_r

What was implemented

• Per-thread accumulation of Σ x² and Σ x·dz with Kahan-style compensation.
• Warp (sub_group) reduction via warp_reduce_sum.
• Cross-warp reduction using local memory (one value per warp) with a single barrier.
• group_broadcast used to distribute inv_r and coeff across the work-group.
• Work-group size: multiple of WARP_SIZE, capped by device limit (≤256), not larger than D.

Optional fast path

• Define GGML_SYCL_RMS_BACK_FAST to disable compensated summation and use plain FP32 accumulation.
• Default remains high-accuracy compensated mode.

---

Validation

Focused tests executed locally:

Test Suite	Result
RMS_NORM_BACK (CPU)	4 / 4 passed
RMS_NORM_BACK (SYCL host/GPU)	4 / 4 passed
Sanity check (NMSE)	≈ 1e-11

Build is warning-free for this code path.

---

Reproduce (build + test)

# Configure & build with SYCL
cmake -B build -DGGML_SYCL=ON && cmake --build build -j"$(nproc)"

# Focused RMS_NORM_BACK tests
./build/bin/test-backend-ops test -o RMS_NORM_BACK -b CPU
SYCL_DEVICE_FILTER=host ./build/bin/test-backend-ops test -o RMS_NORM_BACK -b SYCL0
./build/bin/test-backend-ops test -o RMS_NORM_BACK -b SYCL0

# Optional fast path (less numerically stable)
# add -DGGML_SYCL_RMS_BACK_FAST to your compiler definitions

---

Files Changed (minimal scope only)

File	Purpose
ggml/src/ggml-sycl/norm.cpp	Implementation of ggml_sycl_op_rms_norm_back
ggml/src/ggml-sycl/ggml-sycl.cpp	Operation dispatch registration
ggml/src/ggml-sycl/norm.hpp	Function declaration
docs/ops.md	Mark RMS_NORM_BACK as ✅ for SYCL
docs/ops/SYCL.csv	Mark RMS_NORM_BACK entries as supported

No unrelated files or personal data included.

---

Notes & Risks

• Default path gives high numerical accuracy using compensated FP32 sums.
• Fast path is fully optional (disabled by default).
• Reduction order on GPUs is not bitwise-identical to CPU, but produces NMSE ≈ 1e-11.

---

Reviewers

cc @CISC @NeoZhangJianyu
Looking forward to your feedback. Thanks in advance!

[YaelLogic]

This PR is ready for review.
Tagging @CISC and @NeoZhangJianyu — your feedback would be greatly appreciated whenever you have the chance.
Thanks for your work on maintaining and improving the SYCL backend!

[NeoZhangJianyu]

It's good job!

Thank you!

[NeoZhangJianyu]

@YaelLogic
Please fix the EditorConfig issue.

[YaelLogic]

Hi @NeoZhangJianyu,
the EditorConfig issue has been fixed.
Please let me know if anything else is needed. Thank you.