#2229 sycl: enable vision (--vision cpu) on intel

Descripción

# sycl: wire the CPU image encoder into setup and the server (--vision cpu)

## Summary

The SYCL port's engine already had the full vision path (generate.cpp's --vision: the per-cell (t,h,w) M-RoPE table, GENI requests, SVE1 embeddings files) - only the setup wiring was missing, so docs/INTEL.md listed images as "Not done". This wires the CPU encoder, the same #304 path the AMD port uses. CPU ONLY: there is no GPU encoder build for SYCL (tools/vision knows only -DSTRATA_VISION_CUDA), so --vision yes/gpu are declined with a warning and --vision cpu runs the encoder on the CPU.

## What changed

sycl/setup_intel.py:
- hip_vision returns "cpu" when asked (warns + declines yes/gpu); the forced --vision none is gone;
- to_sycl keeps the config's "vision" section and appends --vision to the engine's args (it dropped both before);
- write_run_script builds the encoder into engine/ with setup.py's own build_vision_cpu when the config carries vision.

sycl/serve/server_intel.py:
- install_embeddings_path: for the strata-sycl.sh layout only (config exe strata-sycl.sh + a sycl_root key) the request's embeddings file is written under the mount and the GENI line names it at /work/...; a deployment where the engine sees the same paths as the server (or any config without both keys) is untouched.

Nothing outside sycl/ changes: the engine, serve/server.py, setup.py and the CUDA/HIP paths are byte-identical to main.

## Extra Notes

Validated on an Arc Pro B70 32 GB, Ubuntu 24.04, engine 0.1.41 (build-sycl-aot, AOT bmg-g31):
- tools/test_setup_sycl.py 17/17
- strata-vision against the pinned llama.cpp (READY 2560; a 672x448 picture -> OK 294 21 14, ~3 s at 4 threads)
- end-to-end test through serve/server_intel.py + the engine (--serve --vision)
- an image_url chat completion returned a correct description of the picture.