attempt to make Appveyor's Cygwin test faster

[Cyan4973]

Cygwin is the longest Appveyor test.

Appveyor is typically the CI which finishes last, and therefore our limiting factor,
presumably because it doesn't parallelize its workload.

Longer term, we should considering transferring some of these tests to github actions,
where parallelization is more effective.

edit : well, this attempt is not greatly effective.
The tests are running noticeably faster, saving > 70 seconds,
but gains are more than offset by extended compilation duration,
likely due to -O1 (instead of "default", aka -O0).
That's unexpected, I was not expecting -O1 to make such a large difference.

edit 2 : tested on my local workstation
compilation of zstd costs about ~x4 more time at -O1 than at -O0.
clang is slightly faster than gcc (about -10%), but it doesn't change the overall picture.

edit 3 : so I scaled back ambitions, and got after the easy wins,
saving approximately 20 sec to Cygwin tests, which is very little.
I'm now wondering if the majority of the delay rather comes from initial package installation.
I see in traces that there are 114 packages being updated before any test actually begins.
That seems like a lot, and is most likely source of substantial delays.
To be investigated, in a separate PR.