[Core] Free KV cache GPU memory on engine shutdown

[markmc]

Related to #24885

Addresses the enable_multiprocessing=False TODO in tests/v1/shutdown/test_delete.py::test_llm_delete

The trickiest part is the single-process case ("inproc" engine and "uniproc" executor") where we can't rely on shutting down child processes to release GPU memory. To address that, we:

1. Call engine_core.shutdown() from LLMEngine.__del__()
2. Free KV cache GPU memory in GPUWorker.shutdown()

Other changes include:

1. Avoid pytest timeout when waiting for GPU cleanup - use the wait_for_gpu_memory_to_clear() timeout parameter to get a nice Memory of devices not free error
2. Print memory usage at start of shutdown tests

[gemini-code-assist]

Code Review

This pull request effectively addresses the GPU memory leak on engine shutdown, especially for the single-process case, by introducing an explicit cleanup path. The refactoring in tests/utils.py to extract check_gpu_memory_usage is a nice improvement for test clarity.

I have two main points of feedback regarding the robustness of the shutdown mechanism:

1. The reliance on __del__ for cleanup in LLMEngine can be unreliable in complex applications with reference cycles.
2. The shutdown method in GPUWorker is now less defensive, which could lead to errors during shutdown if a worker failed to initialize completely.

Details are in the line comments.