If the underlying mechanism's accept_sec_context returns an error, the
spnego accept_sec_context was leaving allocated data in
*context_handle, which is incorrect for the first call according to
Fix this by mirroring some code from the spnego init_sec_context,
which always cleans up the half-constructed context in case of error.
This is allowed (though not encouraged) by RFC 2744 for second and
subsequent calls; since we were already doing it in init_sec_context,
it seems simpler to do that than keep track of whether this is a first
call or not.
Commit By: ghudson