Consider to avoid force_materialization call in to_pandas function

[anmyachev]

force_materialization function takes dataframes from storage, merges them into one and puts them back into storage. The step at which the object is put back into storage can be avoided, resulting in a performance boost. However, this can significantly complicate to_pandas function , so when implementing it someone needs to understand whether the resulting performance gain is worth the increased code complexity. Originally found in #4776 (comment)