Provably Efficient Representation Learning with Tractable Planning in Low-Rank POMDP