[codex] Support latest SGLang integration

[yubofredwang]

Summary

• add a v0.5.12 SGLang Dockerfile and patch for the latest SGLang image family
• use mooncake-transfer-engine-cuda13 in the CUDA 13 SGLang image instead of the generic CUDA 12 Mooncake wheel
• make Mooncake env parsing tolerate Kubernetes service-style tcp://host:port values
• load SGLang before Mooncake in the SGLang integration test to avoid the CUDA 13 pybind import-order segfault

Validation

• docker build -f docker/sglang/v0.5.12/Dockerfile . -t torchspec:sglang-v0.5.12-tgl-cuda13-review-20260529
• docker run --rm --gpus all --entrypoint python3 torchspec:sglang-v0.5.12-tgl-cuda13-review-20260529 ... verified SGLang patch fields and Mooncake import in the safe order
• docker run --rm -v /scratch/ywang/TorchSpec:/work -w /work --entrypoint python3 torchspec:sglang-v0.5.12-tgl-cuda13-review-20260529 -m pytest -q tests/test_mooncake_force_delete.py::TestMooncakeEnvDefaults::test_from_env_accepts_kubernetes_service_port_urls