sycl: fix row calculation when K_QUANTS_PER_ITERATION is 1

[malsbat]

Overview

As noted by in @logari81 in #25063 (comment), multi-turn conversation returned corrupt results after the introduction of #25063.

The root cause is an incorrect row calculation that caused the kernel to overwrite the same row.

Additional issues found while validating this:

• The reordered q5_K kernel only processed one half of each reordered block. Fix the reordered kernel to match the correct unordered kernel.
• The unordered q5_k kernel did not use the K_QUANTS_PER_ITERATION define. This was the only K kernel to not do so.
• The row guard check was incorrect in several of the kernels that could lead to an invalid row access.

Additional information

Validated with test-backend-ops test -o MUL_MAT for permutations of K_QUANTS_PER_ITERATION, GGML_SYCL_ENABLE_OPT, and GGML_SYCL_PRIORITIZE_DMMV.

Manual validation using multi-turn chat to confirm reasonable responses: llama-cli -m /models/Qwen3.6-35B-A3B-Q4.gguf -ngl 99.

Requirements

• I have read and agree with the contributing guidelines: YES
• AI usage disclosure: YES, used to help identify root cause of both the incorrect row calculation and q5_k reordered kernel error.

[arthw]

Could you share the test method to trigger the issue to be fixed?
or
Let's wait for if it can fix issue: #25455

Thank you!

[malsbat]

@arthw, of course:

Prior to the PR:

git checkout 00fa7cb284cbf133fc426733bd64238a3588a33e
# build
cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=ON -DLLAMA_USE_PREBUILT_UI=OFF
cmake --build build --config Release -j$(nproc)

First test, bad output on second prompt:

./build/bin/llama-cli -m /models/Qwen3.6-35B-A3B-Q4.gguf -ngl 99
> how do i bake a cake?
# good response
> customize it for chocolate
# incomplete response

Second test, MUL_MAT_OP:

GGML_SYCL_PRIORITIZE_DMMV=1 ./build/bin/test-backend-ops test -o MUL_MAT
# ...
Failing tests:
  MUL_MAT(type_a=q5_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1)
  MUL_MAT(type_a=q5_K,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1)
  Backend SYCL0: FAIL

After PR, good CLI responses and passing all MUL_MAT_OP tests.

[arthw]

It's good job!

Thank you!