Close the logs-done channels only once

[un-def]

Previously, pullGetHandler closed pullDoneCh on every request it served in the WaitLogsFinished state, and streamJobLogs closed wsDoneCh on every stream that drained after shutdown. Both assumed they run exactly once per server lifetime.

Neither assumption holds. Server.stop() sets WaitLogsFinished as soon as the job is asked to stop, while the server keeps serving until the executor returns -- which does not happen while a process left behind by the job holds the pty slave open, since io.Copy from the master then never sees EOF. The dstack server goes on polling, and the second pull panics:

http: panic serving 127.0.0.1:50670: close of closed channel
...
api.(*Server).pullGetHandler.func1()
    runner/internal/runner/api/http.go:191

/logs_ws has the same defect without needing a stuck executor: nothing limits the number of connections, each is drained by its own goroutine, and two draining together close the channel twice. Both handlers hold only a read lock, so concurrent requests can observe the state together as well.

net/http recovers the panic per connection, so the runner survives it but serves a broken pull endpoint from then on: the job state stops being reported and the run hangs.

Now both channels are closed through sync.Once helpers. The api package had no tests; the ones added here cover the repeated and the concurrent case.

Why the executor fails to return is left alone here.