Allow stderr for git-annex subprocess with records generator

[bpoldrack]

The previous assertion that there's no stderr output to be expected is
wrong. This function is calling git-annex commands, which in turn may
involve arbitrary special remotes. DataLad can not know whether there is
anything to expect on stderr and if such would indicate an error. We
should be able to rely on git-annex returning non-zero.

Closes #7328

[bpoldrack]

FTR: MacOS failures during setup, Ubuntu failures are #7320

[bpoldrack]

@christian-monch : Could you have a look whether that makes sense to you?

Are there any side-effects you can think of, when switching those protocols? Only aim is not failing just b/c there's some output on stderr.

[christian-monch]

@christian-monch : Could you have a look whether that makes sense to you?

Are there any side-effects you can think of, when switching those protocols? Only aim is not failing just b/c there's some output on stderr.

Functionally it should be identical, except from the fact that stderr-output does not lead to an exception.

Having said that, the only reason that the GeneratorAnnexJsonNoStderrProtocol-class was introduced was that the original code raised an exception when stderr-output was received.

[bpoldrack]

Thanks for confirming!

the original code raised an exception when stderr-output was received

Yes, that may well be. But I think it's the wrong thing to do. We are invoking arbitrary annex commands. We cannot possibly know whether any stderr output is expected, especially when annex itself is going to call arbitrary special remote implementations.

[yarikoptic-gitmate]

PR released in 0.18.3