When Dimensionality Hurts: The Role of LLM Embedding Compression for Noisy Regression Tasks