[CI/Test] improve robustness of test by replacing del with context manager (vllm_runner)

[youkaichao]

fixes #5337 .

[youkaichao]

During writing this PR, I find there are already many places where people are fighting against gc, e.g.

vllm/tests/samplers/test_beam_search.py

Lines 41 to 44 in 9fb900f

	# NOTE(woosuk): For some reason, the following GC is required to avoid
	# GPU OOM errors in the following tests using `vllm_runner`.
	gc.collect()
	torch.cuda.empty_cache()

cc @WoosukKwon FYI 😄