execmd: limit the cases where the child process could call the LOGXX functions to really exceptional cases. Previously, the child process could be called with a non-existant command to exec, which would trigger a LOGERR call, which in turn could block because the log mutex had been taken in the father process (and there is nobody in the child to free it). This would manifest itself by 20mn of "selectloop returned 1" messages as the father was waiting for a blocked child until the filter timeout. Other threads would go on, and the timeout would finally trigger, so this did not end up as a failed indexing as long as someone was patient...