Mitigating Privacy Risks in LLM Embeddings from Embedding Inversion