Adds more tests!

[finbarrtimbers]

Here, I add two tests:

1. There was a pre-existing test for ToolUseLLM in open_instruct/tool_utils/tool_vllm.py. Here, I pull this out into test_tool_vllm.py. I also add a mocked vllm instance so that we will automatically run the tool tests as part of our CI. The tests pass on both GPU and CPU.
2. I have updated the beaker integration test so that it runs with 2 vllm engines and uses a smaller Qwen model (0.6B). This test would have caught the timeout issue (example repro). This also succeeds otherwise (example run).