agent/executor: Fix DoSleeper deadlock/goroutine leak

[sharnoff]

Introduced by #371. Classic case of blocking on <-timer.C after !timer.Stop() not always being sound if you've already received from timer.C

Example stack trace:

goroutine 281014 [chan receive, 33 minutes]:
runtime.gopark(0x1?, 0x2?, 0x0?, 0x0?, 0xc00007c000?)
	/usr/local/go/src/runtime/proc.go:381 +0xd6 fp=0xc0017b0cf0 sp=0xc0017b0cd0 pc=0x43d7d6
runtime.chanrecv(0xc001702840, 0x0, 0x1)
	/usr/local/go/src/runtime/chan.go:583 +0x49d fp=0xc0017b0d80 sp=0xc0017b0cf0 pc=0x408a7d
runtime.chanrecv1(0xc001034340?, 0xc0017b0ddc?)
	/usr/local/go/src/runtime/chan.go:442 +0x18 fp=0xc0017b0da8 sp=0xc0017b0d80 pc=0x408578
github.com/neondatabase/autoscaling/pkg/agent/executor.(*ExecutorCore).DoSleeper(0xc001034340, {0x1b971b0, 0xc00149a1e0}, 0x0?)
	/workspace/pkg/agent/executor/exec_sleeper.go:23 +0xd8 fp=0xc0017b0f18 sp=0xc0017b0da8 pc=0x15630b8
github.com/neondatabase/autoscaling/pkg/agent/executor.(*ExecutorCore).DoSleeper-fm({0x1b971b0?, 0xc00149a1e0?}, 0xc002404fa0?)
	<autogenerated>:1 +0x3c fp=0xc0017b0f48 sp=0xc0017b0f18 pc=0x158f1bc
github.com/neondatabase/autoscaling/pkg/agent.(*Runner).spawnBackgroundWorker.func1()
	/workspace/pkg/agent/runner.go:388 +0xe3 fp=0xc0017b0fe0 sp=0xc0017b0f48 pc=0x1577e23
runtime.goexit()
	/usr/local/go/src/runtime/asm_amd64.s:1598 +0x1 fp=0xc0017b0fe8 sp=0xc0017b0fe0 pc=0x471a41
created by github.com/neondatabase/autoscaling/pkg/agent.(*Runner).spawnBackgroundWorker
	/workspace/pkg/agent/runner.go:354 +0x1ec

[sharnoff]

e2e failed because of docker pull rate limit, but the change is pretty low-risk. Merging anyways