Reduce redundancy during the interpreter execution process

Hi, I found some issues when trying to investigate the spark OOM issue:

1. Thread of interpreter executor hasn't been closed when interpreter is restarted;
2. When restart from notebook page, the Zeppelin Daemon are going to close all interpreters, while Zeppelin interpreter server only close the opened interpreters indeed, and it causes in duplicated interpreters kept in server side;
3. After one user starts multi spark sessions, duplicated loaded Classes are founded for the innerScalaInterpreter inside the spark interpreter, since the threads are created in different sub threads, and they don't share loaded class.

I will create a PR to solve the above issues.