fix: Swap the trt_llm_backend base image default off py3-min

[mc-nv]

What does the PR do?

• Dockerfile.triton.trt_llm_backend's BASE_IMAGE default still pointed at
  tritonserver:25.12-py3-min, the image being retired.
• Points it at the published cuda-dl-base equivalent of the same train
  (25.12-cuda13.1-devel-ubuntu24.04, confirmed on NGC) — an image-family swap
  only, no version bump.

Where should the reviewer start?

• dockerfile/Dockerfile.triton.trt_llm_backend line 1.

Related PRs:

• fix: Build on cuda-dl-base devel and ship on inference-runtime server#8994
• fix: Drop the py3-min BASE_IMAGE override from the inferentia SDK build python_backend#458

Related Issues:

• Resolves: TRI-1930

[greptile-apps]

Confidence Score: 4/5

[High risk] Changes the base container image for the inference backend.

The PR is not safe to merge as a default image change until the resulting image includes Triton Inference Server.

Findings

1.  Triton server runtime is missing ▶

Summary

The PR changes the default Docker base from the Triton Server image to a same-train CUDA deep-learning base image. The Dockerfile still relies on its base to supply Triton Server, so the default build no longer produces the documented runnable server image.

_{Reviews (1) · Last reviewed commit: "fix: Swap the trt_llm_backend base image..."}