Support DoFn.setup and DoFn.teardown in Python

DoFn.setup and DoFn.teardown is currently supported in Java but not Python. These would be useful for performing expensive per-thread initialization.

While lazy initialization is a usable workaround, first class support for setup and teardown would encourage consistent conventions and make the API more uniform with the Java version.

This is related to BEAM-3736 (same issue, but for CombineFn).