Use Request::Pause rather than Request::Wait for

Butler::wait_until_finished. Otherwise the following bad
thing happens:

1.  The export code wants to call some Butler functions, so it calls
   calls Butler::wait_until_finished.
2.  This (used to) write Request::Wake into the butler's request pipe.
3.  Imagine that when this happens, the butler is already doing stuff.
4.  Meanwhile, Butler::wait_until_finished is waiting on Butler::paused.
5.  Some time later, the butler finishes its other stuff.
6.  Then it signals "paused".
7.  This causes Butler::wait_until_finished to return, so the export code
   thinks everything's ok and starts calling butler functions.
8.  Then the butler sees the Request::Wake, wakes up, and by unhappy coincidence
   ends up calling read on the same diskstream that the export code has just called.

This causes corruption of the Diskstream buffers, resulting in mantis #4283.

Using Request::Pause instead means that the butler will still wake in step #8,
but should_run will be false, so nothing much will happen and the export code
will be unimpeded.

For future reference, this bug was easiest to track down after adding a debugging
mutex to AudioDiskstream and then try-locking it in AudioDiskstream::_do_refill;
as far as I can see, _do_refill should never be called by two threads at the

git-svn-id: svn://localhost/ardour2/branches/3.0@11163 d708f5d6-7413-0410-9779-e7cbd77b26cf