Can we use LLMs to bootstrap reinforcement learning? -- A case study in digital health behavior change