Learned phenotypic embeddings enable scalable imputation of high-content molecular data elucidating prognostic chromatin signatures