T-FREE: Tokenizer-Free Generative LLMs via Sparse Representations for Memory-Efficient Embeddings