It's up to MLIR folks.
I did this patch as a prototype when I was trying to decouple MLIR compilation passes from hard dependency on NVIDIA GPU driver. The hard dependency has been sufficiently mitigated w/o this patch.
The main benefit is that it makes things more hermetic. Less of a chance that things would go wrong due to the end-user setup (e.g. no space in /tmp, picking up wrong ptxas). Easier to reproduce issues when we need to -- official builds would be guaranteed to use the same PTX compiler as the end user.
If there's a bug in the ptxas compiler, we only need to rebuild with the fixed library. It gives clear idea which version is broken and which is fixed. With ptxas the answer would depend on the specific set up for the individual user, which makes triaging/support much more of a hassle.