viable/strict/1778748069: [ROCm][CI] Stop setting CUDA_VISIBLE_DEVICES on ROCm (#183622)

On ROCm builds, .ci/pytorch/test.sh exported both CUDA_VISIBLE_DEVICES and HIP_VISIBLE_DEVICES for the default test config. The HIP/HSA runtime warns on stderr (agent.cpp:554] Conflicting visibility of agent-N between HIP_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES) whenever both are set and the two filters resolve to a different visible agent, which happens on the multi-GPU CI runners. The warning is benign but leaks into stderr during import torch and breaks dynamo/test_logging.py::LoggingTests::test_logs_out, which compares subprocess stderr against the TORCH_LOGS_OUT file byte-for-byte.

CUDA_VISIBLE_DEVICES is also redundant on a HIP build: torch.cuda._parse_visible_devices already prefers HIP_VISIBLE_DEVICES and ignores CUDA_VISIBLE_DEVICES when it's set.

Set only HIP_VISIBLE_DEVICES on ROCm, only CUDA_VISIBLE_DEVICES otherwise. Matches the existing pattern at line 202 for the distributed config.

Authored by Claude.

Pull Request resolved: #183622

Approved by: https://github.com/atalman