[compute] Correct index permutation in subscribe snapshot optimization

[bkirwi]

#34791 has a bug: it does not correctly permute the data from the arrangement to match the order of columns in the original relation type. The original set of tests did not catch this: I had tests only with default arrangements, which AFAICT include all columns in the original order by default.

• Add a repro of the original bug, and some randomized testing which gives us automatic coverage of this case.
• Fix the bug by applying the appropriate MFP. (Following closely the way the existing sink code makes collections from arrangements.)

Motivation

Previously-unreported bug. (Found during manual testing in staging. 😬)

Tips for reviewer

Here's a nightly that reproduces the issue (before the fix).

[bkirwi]

@def- - Made a small change to the parallel workload in case you want to take a look!

Also, my first Nightly run didn't repro this bug because it was caught by the ignore for this issue. I spot checked a couple builds and didn't find any recent examples; any easy way to check the logs and see if the original issue's cropped up recently?

[def-]

Thanks, test changes lgtm

[def-]

any easy way to check the logs and see if the original issue's cropped up recently?

Easiest way is just removing the ignore and seeing if it still falls over: #34897

We don't currently log expected query errors in parallel-workload, because it would lead to huge amounts of logs.