Two-sample Mendelian randomization: avoiding the downsides of a powerful, widely applicable but potentially fallible technique