fix: make .runpod/tests.json hub tests pass on CUDA 13.0

[TimPietruskyRunPod]

Summary

Three coupled fixes so the re-enabled hub tests from #295 actually pass. Verified end-to-end on a private fork — both basic_inference_test and openai_messages_test green on v0.1.3.

What was broken

After v2.20.0 (#288) bumped the base image to CUDA 13.0, the hub-test flow had three independent failures stacked on top of each other. Each one masked the next:

1. CUDA version mismatch in tests.json — allowedCudaVersions was still ["12.9", "12.8", "12.7", "12.6", "12.5"] so the test pod was scheduled on a driver < 13.0 and container init aborted at the nvidia-container-cli hook with unsatisfied condition: cuda>=13.0.

2. kernels 0.15.1 breaking transformers — kernels v0.15.0 (#544) tightened LayerRepository to require revision= or version=. transformers>=5 still constructs LayerRepository(repo_id=..., layer_name=...) without either (hub_kernels.py:89), so worker import raises ValueError at module-load time. Pinning kernels<0.15 keeps us on 0.14.1 which doesn't have the strict check.

3. Test timeout too tight for vLLM cold start — vLLM 0.20.2 torch.compile + FlashInfer warmup on RTX 4090 takes ~60–70s before the first inference is served (observed: torch.compile took 19.34 s + warmup + model load). The 30s per-test timeout fired before the worker came up, producing context cancelled or timed out: context deadline exceeded for every test even when the worker was healthy. Bumping to 300s.

Changes

• .runpod/tests.json: allowedCudaVersions → ["13.0"]; both test timeout → 300000
• builder/requirements.txt: kernels → kernels<0.15

Validation

Private fork TimPietruskyRunPod/worker-vllm, identical code path, four iterations to isolate each bug:

Tag	Change	Result
v0.1.0	upstream mirror	Container init fail — cuda>=13.0
v0.1.1	+ CUDA 13.0	Worker boot fail — kernels.LayerRepository ValueError
v0.1.2	+ kernels<0.15	vLLM up (torch.compile 19.34s), tests timeout @ 30s
v0.1.3	+ 300s timeout	2/2 tests passing

Refs: DR-1161

Test plan

• PR build kicks off (dev-refs-pull-<N>-merge)
• Hub tests pass on this PR's image: basic_inference_test ✓ and openai_messages_test ✓
• After merge + version bump release, the same tests pass on runpod/worker-v1-vllm:vX.Y.Z