`make check` should use a performance ratchet

[graydon]

We should have performance tests that are "not allowed to get worse" in the testsuite. The operation should involve reading a .json file with the last run values, comparing to current-run values, failing if any regress, and rewriting the file if any advanced. Buildbot can potentially deposit the .json file before each build if it's missing, and harvest it after for display / plotting, and users can just work with the one(s) in their own workspaces.

This requires either objective performance criteria or fuzzy forms of comparison. Prefer objective when possible. valgrind --tool=lackey and perf stat can gather number of instructions retired, which tends to be much more stable than clock time.

Probably easier to base this on metrics gathered from #6810

[huonw]

Triage: We now have some/much of the infrastructure for a ratchet, but I'm not sure if it is enabled on the buildbots. (@brson do you happen to know?)

[brson]

Not 0.8

[steveklabnik]

Triage: no changes here.

[aturon]

Closing for the time being, as this is a pretty open-ended feature request and is unlikely to be prioritized any time soon.