Adaptive Refinement Protocols for Distributed Distribution Estimation under $\ell^p$-Losses