Someone would have to benchmark all those things to quantify the difference :- ). We've seen this quantified with 2990WX where the difference was pretty severe (due to only half of dies having direct memory controller access mostly).
Wendell on youtube had video yesterday where he did this on 3960X and used the Phoronix suite, but that is not exactly useful for us and there wasn't much difference in Vray or Blender.
It affects latency sensitive processes, under which theoretically fall all those you listed. Yet the difference between worst (2133) vs fastest (3600+) is not big at all (5perc.?)with 3rd gen Theadrippers it seems, and is virtually non-existent after 3000.
But most importantly, the price difference between average (2666) and really good (3200) is not big at all, often time the difference is almost nothing. Of course, past this, it quickly exponentially grows expensive, but we can ignore those module since they aren't stable in high capacity anyway. 3200 CL16 is golden standard for Zen3, and doesn't come with any price premium (only price difference between different 3200/16 modules comes down to heatsink and amount of RGBs on, like Corsair LPX vs PRO, or G.Skill RipJaws vs Z-Neo,etc.. no actual performance difference).
Good price right now is 300 Euro for 2x32GB 3200 CL16. You need two of such packages to populate quad-channel, that's 600 Euro for 128GB of fast DDR4 Ram for Threadripper or Ryzen9. In my opinion it is no brainer to buy :- ). In upcoming year, this price might go up. And you can always upgrade to 256GB easily should you ever want to.
Unless you already have existing memory, I would ignore 16GB modules.