images/gpu: update test images for Blackwell (sm_100/sm_120)

[copybara-service]

images/gpu: update test images for Blackwell (sm_100/sm_120)

AI assistance

This change was prepared with AI assistance (Claude Code) and is
labelled Assisted-by: Claude Code in the commit message, per
CONTRIBUTING.md. I have reviewed the change and am able to discuss and
justify it.

Problem

Several GPU test images pin CUDA versions that predate Blackwell, so
they fail on sm_100/sm_120 with no kernel image is available for execution on the device -- under runc exactly as much as under
runsc. These tests therefore cannot pass on Blackwell regardless of
nvproxy.

Changes

• vllm: pinned v0.4.2, whose torch 2.3.0+cu121 compiles for
  sm_50..sm_90. Bumped to v0.27.1 (torch 2.13.0+cu130, built for
  sm_75..sm_90, sm_100, sm_120), pinned by digest.
• nccl-tests: CUDA 12.8 is the first toolkit whose nvcc knows the
  Blackwell arches, and nccl-tests derives its gencode list from the
  toolkit.

The CUDA sample images (cuda-tests, cuda-tests-12-8) already work on
Blackwell unchanged, because the samples embed PTX and the driver
JIT-compiles it forward.

stable-diffusion-xl is deliberately left alone: it installs torch
transitively from the default PyPI index rather than a CUDA-specific one,
so it needs an explicit cu128 index plus a compatible xformers pin.
pytorch is handled in a separate follow-up change.

Verification

On an NVIDIA RTX PRO 6000 Blackwell Server Edition (GB202), the upstream
vLLM v0.27.1 image serves facebook/opt-125m and answers completion
requests under both runc and runsc.

---

🤖 Generated with Claude Code

FUTURE_COPYBARA_INTEGRATE_REVIEW=#14215 from alexhuo2020:blackwell-test-images bec8194