test: backend resource cleanup post-PR #721

[planetf1]

Context

Part of the Testing Infrastructure & Strategy Overhaul epic (item 7). PR #721 (@avinash2692) introduces cleanup_gpu_backend() and directly fixes the core issues: #625 (vLLM redundant servers), #630 (HF memory leaks), #620 (HF metrics OOM).

Objective

Once PR #721 lands, follow up by removing or consolidating workarounds that are no longer needed. Centralise any remaining platform-specific logic into shared fixtures.

Work

• Wait for PR ci: memory management in tests #721 to merge
• Remove or consolidate aggressive_gpu_cleanup() if superseded by cleanup_gpu_backend()
• Remove --isolate-heavy process isolation if no longer needed
• Clean up scattered per-test cleanup code across conftest.py and individual tests
• Centralise remaining platform-specific logic (MPS differences, no-GPU skip paths) into shared fixtures
• Verify CI: backends for vllm testing are redundant and cause CUDA context lock #625, creating multiple LocalHFBackend in pytest caused a memory leak #630, test_huggingface_token_metrics_integration is too heavy #620, test: intermittent failures from vllm tests on LSF cluster #699 are resolved and can be closed

Notes

• Do not start this until PR ci: memory management in tests #721 is merged
• CI: backends for vllm testing are redundant and cause CUDA context lock #625, creating multiple LocalHFBackend in pytest caused a memory leak #630, test_huggingface_token_metrics_integration is too heavy #620, test: intermittent failures from vllm tests on LSF cluster #699 are expected to close with PR ci: memory management in tests #721 — this issue covers any residual cleanup