Unittest NodeTest.LeaderFailWithWitness is not stable.

[ehds]

NodeTest.LeaderFailWithWitness 测试用例：

braft/test/test_node.cpp

Lines 462 to 477 in b37c610

	// apply something when follower
	std::vector<braft::Node*> nodes;
	cluster.followers(&nodes);
	cond.reset(10);
	for (int i = 0; i < 10; i++) {
	butil::IOBuf data;
	char data_buf[128];
	snprintf(data_buf, sizeof(data_buf), "follower apply: %d", i + 1);
	data.append(data_buf);
	braft::Task task;
	task.data = &data;
	task.done = NEW_APPLYCLOSURE(&cond, -1);
	// node 0 is witness;
	nodes[1]->apply(task);
	}
	cond.wait();

这段逻辑应该是 2 个 Replica 和一个 Witness 节点，当选出 leader 后，杀死leader，并用 follower 进行日志提交。
但是在这个 for 循环之前，有可能该 follower 已经成为 leader 了，导致 followers 的个数只有 1 个，使用 nodes[1] 会越界。

@lintanghui

[JayiceZ]

Thanks for your report! We should temporarily disable followers' election before stop the leader to ensure that the raft group will be no-leader state for a period of time deterministically.