Implement import completion cache #861

Steps to reproduce

1. Change current directory to directory containing big sub-tree (ie. system root directory).
2. start bpython executable
3. observe high CPU usage in processor monitor / manager.

Affected versions

0.19 and above

Detailed Description

This problem started after 1a674f0. It seems that new find modules implementation based on importlib.machinery.FileFinder is much slower than previous one.

Should bpython traverse whole sub-tree on each startup? Quite often I do start it from my home directory and traversing all it's contents including mounted remote filesystems seems wasteful.