runtime: scavenger's frequent wake-ups interfere with runnext

[bcmills]

From windows-amd64-longtest (https://build.golang.org/log/3fbe37b41ed88143f4b6d6f6e277e99891537be7):

--- FAIL: TestPingPongHog (0.11s)
    proc_test.go:460: want hogCount/lightCount in [0.2, 5]; got 51000000/6733000 = 7.574632407544928
FAIL
FAIL	runtime	288.181s

CC @aclements @mknyszek

[bcmills]

linux-amd64-longtest: https://build.golang.org/log/8f4ed420d92d7a3872a0701928f301386266e77f

I'm guessing this probably turned flaky due to the goroutine-preemption changes?

[bcmills]

This is preventing me from getting meaningful TryBot results on Windows for CL 204521.

(https://storage.googleapis.com/go-build-log/66ca91a7/windows-amd64-longtest_cfa17417.log)

[mknyszek]

This is now failing much more consistently on the linux longtests, as of 21445b0.

The scavenger is likely going to sleep and waking up much more often than before (though I'm not sure that it's enough to still matter. The time spent "measuring" is 100 ms, though, so I suppose I could see that making this worse.

[gopherbot]

Change https://golang.org/cl/206098 mentions this issue: runtime: skip TestPingPongHog on builders

[mknyszek]

I'll take a look into this since Austin is currently focused on the memory corruption issues.