Add an asynchronous version of _read_all().

This API is not fully consistent with the normal expectations of a
non-asynchronous version.  Consistency between the sync and async version is
probably more important.

The API will still bind correctly, but access to all functionality will
not be available: specifically, in the case of an error, higher level
languages will be unable to determine how many bytes were successfully
read before the error.  Most users will probably not want to use this
information anyway, so this is OK -- and if they do need the
information, then they can just write the loop for themselves.

Heavily based on a patch from Ignacio Casal Quinteiro.

https://bugzilla.gnome.org/show_bug.cgi?id=737451