cAdvisor constantly polls cgroup data for metrics that have been disabled

[eero-t]

This is continuation on #2897 issue. After that, cAdvisor code was change to support enabling/disabling all metrics, including "cpu" & "app" ones in #2910.

Setup:

• Ubuntu 20.04 host (= cgroup v1)
• Few days old git head of runC & cAdvisor (and drm-tip kernel)

Use-case:

• Start cAdvisor with only one metric enabled: cadvisor -enable_metrics hugetlb

Expected output:

• Only the enabled metrics are queried (in this case, "hugetlb")

Actual output:

• cAdvisor constantly queries "cpu" & "pids" cgroups

These are the opened file & dir names from 1 min of tracing:

# strace -f -e openat -p $(pidof cadvisor) 2> trace.txt
^C
# grep openat trace.txt | grep -v "openat resumed" | cut -d'"' -f2 | sed 's%^.*/%%' | sort | uniq -c | sort -nr
   2002 cpu.shares
   2002 cpu.cfs_quota_us
   2002 cpu.cfs_period_us
   1976 pids.max
    134 system.slice
...

[eero-t]

@Creatone does this look as OK fix for this: eero-t@e57ee46

At least in my own testing it fixed the redundant polling.

[Creatone]

@eero-t Could you prepare PR :)?

[eero-t]

@Creatone #2940