Allow negative stride(A,2) in gemv!

[sostock]

This replaces #42012 as another attempt at fixing #41513 (comment). It implements the idea from #41513 (comment).

Instead of just allowing stride(A,2) < 1 in trivial cases, the requirement stride(A,2) ≥ max(1, size(A,1)) is relaxed to abs(stride(A,2)) ≥ size(A,1) || size(A,2) ≤ 1.

We could also think about relaxing the requirement stride(A,1) == 1 to abs(stride(A,1)) == 1.

Closes #42012.

[sostock]

Since #41513 is now marked for backport, this PR (which I prefer over #42012) should be backported as well (as discussed on discourse).

Closing and re-opening to run the tests against current master.

[dkarrasch]

IIUC you need to rebase manually. Closing and re-opening does not rebase onto current master. I think we should get this one merged (and then backported), and then put #42957 on top.

@N5N3 would you be available to take a quick look?

[N5N3]

LGTM, the only (little) improvement I can found is simplify the pointer calculation of A (L679).

[dkarrasch]

LGTM, the only (little) improvement I can found is simplify the pointer calculation of A (L679).

Thanks for the speedy response. Since this PR is supposed to be backported, I don't think we should mix things here, but keep it as is, backport, and develop further on master/v1.8 in your PR.

[dkarrasch]

I believe the test failures are nothing but internet connection failure at the time. I'll rerun and merge once tests (essentially) pass.

[sostock]

I just rebased onto the current master. The LinearAlgebra/blas tests pass on my machine.

[KristofferC]

This causes a bunch of packages to fail on 1.6: https://s3.amazonaws.com/julialang-reports/nanosoldier/pkgeval/by_hash/c7daef8_vs_7be0dcd/report.html. So I will drop this from the 1.6 backports.

[sostock]

Backporting #41513 but not either this or #42012 might also break things (see #41513 (comment)). I will investigate what happens here.

Maybe we could backport #42012 instead?