Do not register it if keys are disabled. Do not leave it dangling if
thread creation fails.

No need to duplicate debug from net_Connect()...

p_finder is not visible to other threads at this point.

From the downloader thread, b_close is always true.

vlc_cancel() is not used, so this was a no-op.

b_error can only become true after the download thread is joined, after
which point it is never evaluated. So we can treat it as always false.

The code size saving in vlc_cleanup_run() is marginal and premature
optimization. In practice, vlc_cleanup_run() makes the source code
harder to follow/read, confuses static analyzers and generates false
positive clobber warnings (on some OSes due to long jumps).

It did exercise some of the cleanup code paths though.

Cancellation cannot occur here.

If the input state were initially "smaller or equal to" PAUSE_S, the
loop would just run forever and deadlock. So assume it is always larger
and remove the slack.