From @rjzamora in #8899 (comment):
but my preference moving forward may be to deprecate gather_statistics as a public option. Right now, the only time the user should ever specify gather_statistics, is when they are doing it to insist on gather_statistics=False. For the case you are highlighting here, the suggested argument is split_row_groups=True (which doesn't actually require any row-group statistics to be gathered, but does require the parquet metadata to be parsed, as you said).