Optimize Delaylines: block-process (not one sample at a time)

Optimize Delaylines: block-process (not one sample at a time)

This also tweaks fade behavior when the latency changes to prefer a
x-fade when possible.

This new variant does not support concurrent re-allocation and
execution. Hence the auto-connect thread needs to take a lock before
updating latencies (actually there's no need for an explicit update with
built-in backends, so this case remains to be updated further)