Save some work by only tailing it once at startup; if we glob it, then we have
to check every time and we then end up having to also do the work in the
logstream of handling the "already closed".

The work of globbing a single pattern is separate from iterating over all
patterns, setting us up to start a goroutine per pattern instead of an
iterator.

Issue: #331

We would like to use this in per-pattern goroutines.

Issue: #331

Instead of collecting sockets and tailing them after, just tail them straight
away.

This makes the "early exit" check for "no things" broken, so replace that with
a check in the final shutdown handler that there were no active globs.  This
will be replaced in the next few changes once migrated to goroutines and we can
use only the `WaitGroup` for counting globs, too, but in the meantime this is
more correct than the previous code.

Issue: #331

Restructure the testserver constructor.

We get greater control of what's being polled for by doing so.  This also
removes noise so in a few changes time I can clean up the log pattern poller.

…eams.

This gives us greater control in test.

This has no effect on test behaviour.

…tter.

We now can explicitly control these two events, making the tests faster.

Update the test tail comment explaining what the waker count parameters do; I
keep forgetting, this hint helps!

This lets us await a goroutine per pattern, and if there are none, no waiting
need be done, making shutdown cleaner.

The completion poller is still a bit of a mess though.

This is not a routine we wait for completion of to shut down the tailer, it's
shut down when the tailer is exiting.

Move the stdin check to logstream so it can use it too.

Rename the waker option and the goroutine launcher as it's not just for stale
logs anymore.

Fix tests that were relying on the pattern poller for log completion.

We pass the context to the functions so we don't need to keep it in here.

…the directory.