On the Learnability of Knot Invariants: Representation, Predictability, and Neural Similarity