Version rolling is not possible with most (stratum V1) pools; the ticket send back to the pool does not include nversion.
Nearly every stratum pool I have seen supports the version-rolling extension. that adds a version_bits field to mining.submit. afaik It's necessary to support ASICboost.
ntime is no problem though, but you 'only' gain 60x with that, so you will only have to send work to a single asic @250GHs once for every work from the pool, or even @500GHs on ck's pool (as the work time is now reduced from 60s to 30s).
Where does 60x come from? every 30s is pretty leisurely compared to every 8ms @ 500 GH/s with nonce rolling alone!
Some sort of block generation will be unavoidable sooner or later unless they do something with the serial interface to make it fast enough (a normal T17 needs a ~1.5Mbaud connection between the asics, so a S19j @120GHs would need roughly 3 times that speed without block generation).
120 TH/s over 3 hashboards = 40 TH/s each, which means each hashboard can cover the 32 bit nonce space in 0.1ms. The newer ASICs have more midstates too, so that means bigger job packet. No way serial can keep up.