flake TestFreezeBeforeSet on CentOS 7 and 8

Happened once in Cirrus CI (https://cirrus-ci.com/task/5418288451682304?logs=unit_tests#L549) while running CI for https://github.com/opencontainers/runc/pull/3512

=== RUN   TestFreezeBeforeSet
=== RUN   TestFreezeBeforeSet/slice,skip-devices
    freeze_test.go:171: failed to write "0\n": write /sys/fs/cgroup/cpuset/system.slice/system-runc_test_freeze_1.slice/cpuset.mems: no such device
=== RUN   TestFreezeBeforeSet/scope,skip-devices
=== RUN   TestFreezeBeforeSet/slice,will-freeze
=== RUN   TestFreezeBeforeSet/slice,pre-frozen,will-freeze
=== RUN   TestFreezeBeforeSet/scope,pre-frozen,skip-devices
=== RUN   TestFreezeBeforeSet/scope,pre-frozen
--- FAIL: TestFreezeBeforeSet (0.41s)
    --- FAIL: TestFreezeBeforeSet/slice,skip-devices (0.06s)
    --- PASS: TestFreezeBeforeSet/scope,skip-devices (0.07s)
    --- PASS: TestFreezeBeforeSet/slice,will-freeze (0.07s)
    --- PASS: TestFreezeBeforeSet/slice,pre-frozen,will-freeze (0.06s)
    --- PASS: TestFreezeBeforeSet/scope,pre-frozen,skip-devices (0.09s)
    --- PASS: TestFreezeBeforeSet/scope,pre-frozen (0.05s)

I've seen similar bugs but not this one. Filing just to ensure it's not lost.

Start with TestFreezeBeforeSet in freeze_test.go, especially the failure at line 171, and review the Cirrus CI task linked in the issue. Reproduce the failure on CentOS 7 or 8 and compare the slice and scope cases. Done means the intermittent cpuset.mems failure is explained and the test or its environment handling is reliably corrected.