Update perf to possibly also accomodate runners with debian:bullseye.

The output of perf seems to be slightly different on runners with different Debian versions. Compare a job run by the Radboud runner vs a job run by the MPI runner.

Notice that the per file statistics from perf talk about instructions:u and cycles:u in the RefinedC job, whereas in Diaframe it is just instructions and cycles.

Consequently, for Diaframe the curl request to coq-speed complains about an invalid syntax. I tried manually uploading the build-times, replacing instructions with instructions:u and then coq-speed no longer complains.

It seems to me that specifying -e instructions,cycles actually results in the intended -e instructions:u,cycles:u on the MPI runner, but not on the Radboud Runner. This change should fix that and be compatible with both.