This guy Keith Howard seems to have done it 10 years ago:
https://www.moon-audio.com/mwdownloads/download/link/id/159/
So, 300x real time using a 10 year old single core desktop computer, not using a GPU, and most likely not using speed optimized code. That 300x would be smoked using a current GPU card. And GPU cards have many parallel processors/cores, just like the FPGA.
Even if not real time it is fine with me, it can take its time, it only has to be done once.
Size is no problem either: 1000 CDs = 600GB. x16 = 10TB, easy for a 12TB drive (€250). I have only about 300 CDs, so about 3TB only.
But the big question is how does it compare to MScaler SQ wise. Keith suggests it could be even better than MScaler, because it does not need simplifications and optimizations, just brute force to do the ideal process.
Another thing he says is that programming the software is much simpler than the FPGA (which I think we can all agree), moreover because the FPGA real time operation requires the genius Rob optimizations.
I am just pointing this article. I am not doing any judging. But I would make two points:
Firstly, I agree with you, why doesn't anybody do it? I am not convinced by HQ Player from what I read (I have not tried it yet because it does not run on Windows 7).
Secondly, I suspect that besides Rob optimizations that optimize speed and are only necessary for real time as in MScaler, Rob filter likely has other details that are crucial for the SQ result (independently of speed) and that come from all his years of research and experience, and specially his genius. But obviously I do not have the knowledge to judge it.