Add Arc B390 iGPU (SYCL) benchmark

[hungho77]

Adds the Arc B390 iGPU under ggml's SYCL backend to the Core Ultra X7 358H report, and a row for it in the summary table in docs/benchmark/README.md.

• Same protocol as the other reports: 3 warmups + 20 timed reps per process, best of 3 processes, then the flag screen and a full re-run of the fastest flag set.
• Measured at 1adf078, llama.cpp b11223, oneAPI 2026.1, GPU compute runtime 26.35.39758.10. The section has its own setup table, because the existing one describes the OpenVINO run at 450992c.
• SYCL runs through the OpenCL adapter (UR_ADAPTERS_FORCE_LOAD=.../libur_adapter_opencl.so.0). The Level Zero adapter segfaults with this machine's loader (1.16.1).
• 11 archs measured. BitVLA is N/A off CUDA. OpenVLA-OFT was not run; the summary table marks it n/r, a new marker defined in the legend.

p50 at the defaults (ms):

Model	Arc B390 (SYCL)	Arc A380 (SYCL)
Octo-Small	17.1	61.0
TurboVLA	45.7	103.6
VLA-JEPA	104.8	356.8
VLA-Adapter	145.0	410.2
GR00T N1.5	156.8	472.7
GR00T N1.6	183.7	647.9
GR00T N1.7	184.9	618.2
SmolVLA	228.2	736.5
π0	367.5	924.1
π0.5	369.9	948.2
Evo-1	536.5	1048.8

The A380 numbers are from a different GPU, commit, driver and Level Zero adapter, so this is a side-by-side, not a re-check of arc-a380.md.

Notes in the report:

• GR00T N1.6, GR00T N1.7 and VLA-JEPA settle at one of two speeds per process, 8–9% apart. For these three, flag differences in the screen are within that noise.
• The 1adf078 build links vla_core with -fsycl-device-code-split=per_kernel and -cl-fp32-correctly-rounded-divide-sqrt (from PR FoldQuant W8A8/W4A4 inference and FoldQuantVLA quantized-model converter #33). These link options also reach ggml-sycl's kernels, so this build is not the same binary as a SYCL build at c93ca0a.