feat(dtk): post-operation to pin numpy and remove nixl ep for vllm 0.18.1

[thxCode]

Summary

Add post-operation 20260901_dtk_pin_numpy_remove_nixl_ep to patch the released gpustack/runner:dtk26.04-vllm0.18.1 image in place, backporting #265 (commit range bc1428f..372b411) to the already-published tag:

• Pin numpy==1.26.4 (uv pip install --no-deps) with a build-time version assertion — the prebuilt DTK Torch is compiled against the NumPy 1.x ABI, and NumPy 2.x breaks distributed init at Tensor.numpy().
• Remove CUDA-only NIXL EP distributions (uv pip uninstall nixl nixl-cu12 nixl-cu13 || true) with a build-time assertion that nixl_ep* modules are no longer discoverable — they make vLLM load an extension requiring NVIDIA libcuda.so.1, which fails on Hygon DCU.

Package layout follows existing post-operation conventions (matrix.yaml + dtk/Dockerfile), and the operation is logged in pack/.post_operation/README.md.

Also ignores the local .claude/ tooling directory.

Validation

• pre-commit run on all changed files — green.
• Local matrix expansion smoke — exactly one job: dtk26.04-vllm0.18.1-linux-amd64.
• Pack workflow dry-run on this branch — green: https://github.com/gpustack/runner/actions/runs/33489120235

Rollout (after merge)

1. Pack workflow: post_operation=20260901_dtk_pin_numpy_remove_nixl_ep, backend=dtk, target=vllm, tag=dtk26.04-vllm0.18.1, dry_run=false, for_release=false → pushes -dev tag.
2. Runtime validation on Hygon DCU with the -dev tag.
3. Re-run with for_release=true to overwrite the released tag, then check off the README entry.

[gemini-code-assist]

Code Review

This pull request introduces a post-operation patch for the Hygon DTK 26.04 backend running vLLM 0.18.1. Specifically, it adds a Dockerfile that pins numpy to version 1.26.4 to restore NumPy 1.x ABI compatibility with Torch, and uninstalls CUDA-only NIXL EP packages to prevent execution failures on Hygon DCU systems. It also configures the build matrix rule for this patch, updates the .gitignore to exclude local Claude agent tooling, and documents the new task in the README. There are no review comments, so no additional feedback is provided.