[Bug]: Set LD_LIBRARY_PATH to include the 'standard' CUDA location

[smarterclayton]

Purpose

NVIDIA changed the CUDA library install path in their base
images in 12.8, but some container providers automount drivers
in the old location. After discussion with NVIDIA future CUDA
images (13.0+) will revert to the old path. Until then, reduce
the friction for new vLLM users running on GKE by including
the old / future mount path ahead of the built in CUDA to
allow automount logic to work out of the box.

Fixes #18859

Test Plan

Verify the built image works with an automounted CUDA version
provided by the container platform at /usr/local/nvidia/lib64.
Start vLLM in a minimal config on GKE with automounted drivers.

Test Result

In progress ...

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request aims to improve compatibility with container platforms like GKE that automount CUDA drivers in the /usr/local/nvidia/lib64 path. The change prepends this path to LD_LIBRARY_PATH to ensure that the automounted drivers are prioritized.

My review identifies a potential security vulnerability in how LD_LIBRARY_PATH is modified. If the existing LD_LIBRARY_PATH is empty, the change could lead to including the current directory in the library search path, which is a security risk. I've provided a suggestion to fix this by conditionally adding the path separator. This ensures the Docker image remains secure while achieving the desired compatibility.