sycl: explicit-SIMD decode kernels for PTQ1_0 and PQ2_0 (1.3x PQ2_0 decode)

[dwymark]

Overview

On Intel GPUs, single-token decode of PTQ1_0 and PQ2_0 runs through MMVQ, where the sub-group kernels from PrismML-Eng#235 set the decode speed. This PR adds explicit-SIMD (ESIMD) MMVQ kernels for both formats: each work-item takes one weight row, spreads the row's blocks over 16 lanes, and accumulates q8_1 dot products with dp4a, four rows per work-group. They compile only under the Intel compiler and follow a new GGML_SYCL_ENABLE_ESIMD switch, default 1; setting it to 0 restores the sub-group kernels.

It also turns a one-row, single-index F32 GET_ROWS into a vectorized copy. The recurrent-state layers of the qwen35 graph issue that shape once per layer per token.

Prompt processing is unchanged: batches above the MMVQ width take the FP16 GEMM from PrismML-Eng#278.

Additional information

Intel Arc 140V (Lunar Lake iGPU), Windows, oneAPI 2025.3, driver 32.0.101.9030, Release. Base prism @ 88c4bc6, Ternary-Bonsai-2-27B.

llama-bench -ngl 99 -fa 1 -p 512 -n 128 -r 1, four rounds in separate processes 30 s apart with the arm order flipped each round; median (range) in t/s:

	base	this PR
PQ2_0 tg128	6.79 (6.70-6.84)	8.88 (8.80-8.98)	1.31x
PTQ1_0 tg128	5.94 (5.93-5.95)	7.03 (7.02-7.04)	1.18x
PQ2_0 pp512	139.3 (129.2-142.1)	135.4 (122.4-142.0)
PTQ1_0 pp512	147.8 (143.6-148.8)	147.6 (104.4-149.2)

With GGML_SYCL_ENABLE_ESIMD=0, this PR decodes PTQ1_0 at 6.07 and 6.08 t/s, against 6.99 with it on. This laptop lowers its memory clock under sustained load, which accounts for the spread in prompt speed.

Correctness:

• test-backend-ops -o MUL_MAT: PQ2_0 216/216 and PTQ1_0 167/167 on base and this PR, with oneDNN, with GGML_SYCL_ENABLE_DNN=0, and with GGML_SYCL_ENABLE_ESIMD=0. 48 cases per type have n = 1 at the model's row lengths and take the new kernels. -o GET_ROWS -p type=f32: 11/11.
• Decode logits, full vocabulary, at 5 positions over 512 prompt tokens and 32 decode tokens. PTQ1_0 is bit-identical to base. PQ2_0 against base: normalized MSE 3.0e-5, mean KL 9.9e-5, max KL 2.3e-4, same top token at every position. Against a CPU-backend reference, both builds sit at the same distance (PQ2_0 normalized MSE 1.00e-3 for this PR and 0.99e-3 for base). With GGML_SYCL_ENABLE_ESIMD=0 both formats match base exactly, so the GET_ROWS copy changes nothing.
• llama-perplexity --kl-divergence, WikiText-2, -c 2048 -b 512, 8 chunks, against base: mean KLD 0.000000, max KLD 0.000056, same top 100% for both formats. These batches take the GEMM path, so this mainly confirms prompt processing is untouched.

Only tested on Xe2.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Claude Code (Anthropic) wrote the kernels while profiling decode, rebased them onto current prism, and ran the measurements above on my machine, under my direction. I reviewed the change and can explain it.