Fix potential connection storm

[wenhaocs]

This affects 5.2.0.

NebulaPool creates multiple clients. When strictly_server_healthy is set to false, each client creates a connection to every server. Consider a scenario where one Graphd instance goes down. When calling get_client, if test_on_borrow is set to true and ping_before_execute is set to false, the client will send a ping to one host in a round-robin manner. If the ping happens to be sent to the failed host, the ping will fail, and the client will be evicted, causing all of that client’s connections to all hosts to be closed. Since the ping target is effectively random, it is possible that all clients are evicted because they ping the failed server. When the number of clients in the pool drops below a threshold, the pool will recreate clients. However, the newly created clients may again be closed due to ping failures. This entire process can lead to a connection storm.

[Nicole00]

After the refactoring:

1. strictly_server_healthy only affects the pool's initialization behavior. When set strictly_server_healthy to true, the pool can only be initialized successfully when all Graphd instances are available.

2. Each client maintains only one connection. When test_on_borrow is set to true, if a Graphd instance goes down, the corresponding client will be destroyed and removed from the pool. Users will then obtain a usable client. When the number of clients in the pool falls below the threshold, new clients will be created. During client creation, the system will poll multiple Graphd addresses until a usable client is successfully established.