net/http/transport: fixed wait time in TestTransportPersistConnLeak is too short

[millerresearch]

The plan9_arm builder sometimes fails in net/http/transport tests with output like:

 --- FAIL: TestTransportPersistConnLeak
     transport_test.go:1032: goroutine growth: 59 -> 162 -> 66 (delta: 7)
     transport_test.go:1033: too many new goroutines

This test has a fixed sleep time of 100ms to allow goroutines to exit after CloseIdleConnections is called. On a Raspberry Pi running many simultaneous tests, this is not always long enough.

A similar function TestTransportPersistConnLeakShortBody has a sleep time of 400ms for the same purpose (originally 50ms but revised upwards in CL 12960046). This seems to be plenty of time for plan9_arm, but on faster hosts will usually be much more than necessary.

Instead of just increasing the sleep time, I propose replacing the fixed sleep in both tests by calls to a new function which will do shorter sleeps in a loop until the number of goroutines reaches the desired threshold or an overall timeout expires.

[gopherbot]

CL https://golang.org/cl/20922 mentions this issue.