single pass scan group virtualization

[sortraev]

This PR adds group virtualization to SegScan.SinglePass. The generated single pass scan code now respects the suggested/requested num_groups/block_size, which in turn allows scanomaps with array construction in the map KernelBody, since mem expansion bases its expansion on this information.

Also

• status flags (for the lookback step) are now initialized inside the kernel
• some light code refactoring

[athas]

Please also add a comment somewhere in the module (it can be in the header) outlining the virtualisation strategy.

[athas]

There is a style issue, but worse, it also appears some of the scan benchmarks fail with the CUDA backend. (There's also a bunch of other things that fail because we are moving servers - you'll have to pick out the interesting failures from the wreckage.)

[sortraev]

There is a style issue, but worse, it also appears some of the scan benchmarks fail with the CUDA backend. (There's also a bunch of other things that fail because we are moving servers - you'll have to pick out the interesting failures from the wreckage.)

Yes, I see now that the benchmarks also fail when I run manually on the A100. I think the problem had to do with our change of the status flags initialization (we simply changed it to align with Cosmin's and our own prototype), but I'm not exactly sure why -- it was not a problem on our own 4090.

Anyway, I have pushed a commit which passes the benchmarks when I run them manually on the A100, so hopefully the bug has been squished (assuming this was the bug I was looking for; admittedly I don't see why it would be a bug). I will look into the other change requests and CI failures at a later time -- thanks for the comments