This fixes #2173.

In this PR:

• By default, worker stdout/stderr on all machines (including the single machine case) are written to files and streamed to the driver via file on disk -> log_monitor -> redis -> driver.
• If the log monitor dies, an exception is pushed to the driver.
• We now shutdown driver threads when we call ray.shutdown(). This is most relevant for our tests.

Possible TODO:

• Make sure we don't ignore files once a large number of workers have been created.
• Think through relevant thread safety issues.
• Format print statements to include node and process ID of process that logged the statement, maybe function name as well.
• Code cleanups (hard coded strings).
• Isolation between drivers, ideally only print logging statements for the relevant driver.
• Think through the API (naming, and in particular, the interaction with redirect_output and redirect_worker_output).
• Tests
• Push error to all drivers if log monitor fails.
• Test log monitor failure.