Content-Style Learning from Unaligned Domains: Identifiability under Unknown Latent Dimensions