[codex] migrate GPU stack to CUDA 13

[bradhilton]

Summary

• migrate ART GPU dependencies, lockfiles, and image config from CUDA 12.8/cu12 to CUDA 13/cu13
• move Torch/TorchVision/torchao to cu130 wheels and TransformerEngine to transformer-engine-cu13
• update vLLM runtime pins away from cu128/cu129/cu12 and rebuild CUDA source extensions in the CUDA 13 image
• update CI/cache/dev image defaults and CUDA/CUDNN/NCCL include/library paths for CUDA 13

Validation

Validated on CoreWeave ext-collab2 node g122252:

• GPU: NVIDIA H200
• Driver: 595.71.05
• nvidia-smi CUDA: 13.2
• nvcc: CUDA 13.0
• NCore image tag: 2.45.0

Commands/checks run:

• uv lock
• cd vllm_runtime && uv lock
• uv lock --check && (cd vllm_runtime && uv lock --check)
• uv sync --extra backend --extra megatron --extra tinker
• cd vllm_runtime && uv sync --no-dev
• core CUDA import smoke for torch==2.11.0+cu130, CUDA 13.0, and H200 visibility
• extension import smoke for torchao, transformer_engine.pytorch, apex, deep_ep_cpp, quack.layout_utils, and PyTorch flex attention
• ldd sanity on 14 CUDA shared objects across Apex, DeepEP, TransformerEngine, and torchao; no not found, including no missing libcudart.so.13
• focused Megatron tests: 21 passed around packed/flattened attention, compile flags, and provider support
• tiny Kubernetes dev smoke: uv run --extra backend --extra megatron dev/yes-no-maybe-megatron.py

Final image built and prewarmed:

• docker.io/bradhiltonnw/art-gpu:cuda13-codex-20260612-122156
• digest: sha256:009d5aa9cbe46a5a2855d44adc6bc3c272b268eb264bccc77e0d247f1b4ce7f0

Notes

• Remaining non-fatal runtime warnings are documented: xFormers is still pulled by Unsloth and its current wheel warns about a cu128 build; this path is not used by the Megatron/flex-attn smoke.
• The tiny dev smoke reached train/eval but skipped gradient updates because one rollout produced no advantage signal.