Handle hdfs_namenode_principal properly and refactor auto configuration logic

[garthcn]

This is related to this issue: #175

See this comment: #175 (comment)

This might not be the ideal way since removing the side effects here and here might be a better solution.

Let me know if this pull request makes sense and really appreciate your great work on the library!

[garthcn]

More detailed description:

The problem: #175 (comment)

Changes in the PR:

1. Remove class variables in HDFSConfig. Calling get_external_config() will return a dict like this:

   {
       namenodes: [{'namenodes': <host>, 'port': 1234}],
       use_trash: True,
       use_sasl: True,
       hdfs_namenode_principal: <principal>
   }
   

2. In places where the class variables in HDFSConfig were referenced, call get_external_config() and use the values returned in the dict.

3. In Client, HAClient, add the option to pass in hdfs_namenode_principal. AutoConfigClient will call get_external_config() and pass hdfs_namenode_principal to HAClient.

4. In SocketRpcChannel, SaslRpcClient and RpcService, change __init__ to accept the hdfs_namenode_principal which will be passed down the chain from Client or HAClient.

5. Update test cases to reflect the change

[wouterdebie]

I'll talk to @ravwojdyla tomorrow, but LGTM for now.

[wouterdebie]

Would you be able to squash this commit?

[garthcn]

@wouterdebie I just squashed the commits into one. Let me know if you have any comments/feedback. Thanks!

[ravwojdyla]

We eventually want to refactor the way we handle config/ctor, but that is out of the scope of this bug/PR. Thanks @garthcn for great investigation and patch. lgtm.