[XPU] feat: add Docker image for verl + Intel XPU plugin

[kahlun]

Summary

Adds a Docker image for running verl + this plugin together on Intel GPU.
Builds this repo (verl-hardware-plugin) as the container's build context and
clones verl-core in at a pinned ref, rather than the other way round. See
docker/intel_gpu/README.md for build/run instructions.

Motivation

No vendor has shipped a Dockerfile in this repo before. verl-core's own
docker/rocm and docker/ascend use verl-core as the build context — that
shape doesn't fit a plugin repo, where the natural build context is the
plugin itself and a known-good verl-core is something you pull in, not
something you're developing against.

Changes

• docker/intel_gpu/Dockerfile.intel_gpu (new) — Intel DLE 2026.1.0 base,
  Intel GPU user-mode driver install, vLLM built from source (pinned
  v0.27.0) for XPU, verl-core cloned at VERL_GIT_REPO/VERL_REF, this
  plugin installed --no-deps -e . from the local build context on top.
• docker/intel_gpu/requirements-intel-gpu.txt (new) — verl's own runtime
  deps for this image (torch==2.13.0+xpu hard pin, transformers>=5.17.0
  for the flash-attn2 Hub-kernel fallback — see the file for why that
  diverges safely from verl's project-wide transformers pin).
• docker/intel_gpu/README.md (new) — build/run instructions, dependency
  table, known Level-Zero-IPC workarounds for multi-GPU.

One decision a reviewer is likely to question:

VERL_GIT_REPO/VERL_REF default to kahlun/verl:intel-xpu-plugin-e2e-vllm,
not verl-project/verl:main. This plugin's platform_xpu.py/profilers/
override four PlatformBase hooks (is_reduce_avg_supported,
attention_utils_module, profiler_markers, dist_profiler_cls) that don't
exist on verl-project/verl:main yet — they're in an unmerged upstream PR
(verl-project/verl#7917). Building against main today would silently no-op
those overrides. Switch both build args once #7917 merges.

Testing

• docker build -t verl-intel-gpu:latest -f docker/intel_gpu/Dockerfile.intel_gpu .
  — not yet run in this environment; flagging rather than claiming green.
• Runtime sanity checks from the README (torch.xpu.is_available(),
  vLLM platform, verl platform/vendor) after build.
• End-to-end GRPO run — tracked separately as the e2e PR, not blocking
  this one.

Acceptance Baseline (for new hardware adaptation PRs)

Skipped — this PR adds a build artifact, not a platform/engine change.

Checklist

• Code follows the project's style and passes pre-commit checks — not
  yet run against this diff.
• Documentation updated — docker/intel_gpu/README.md added alongside.
• No secrets or credentials included