Move its code to DeferEvent instead.

Allows defining a list of supported audio formats, and allows
switching on and off DoP with certain formats.

This is a first rough draft.  The setting syntax and its semantics may
still be redesigned.

There is no documentation on whether calling shout_metadata_add()
multiple times on one instance is allowed.  To be sure, let's allocate
the object on demand each time in SendTag().

This is more robust, for example it allows using the Shine encoder
plugin instead of LAME.

Passing it by value is actually smaller (32 bit) than the rvalue
reference (64 bit pointer), and it ensures that the object is consumed
after the call returns, no matter how the methods are implemented.

Migrate callers to GetSocket().Read(), GetSocket.Write(), which is the
same.

Maybe this was once a required kludge for Windows, but it's not
anymore.