detach the managed job controller from job submission

[cg505]

Previously, the ray driver program as well as a ray worker stayed in use for the entire runtime of a managed job. Now, the job controller will detach from the submitted job/ray driver and continue running in the background.

This means we have to manually manage logging as well as liveness of the controller process. Two new directories are introduced for this purpose as well as plumbing.

• TODO: limit the parallelism of the controller process (high limit, something like 3x GB of memory) and the parallelism of sky launch (more limited, like 1x GB of memory/4x CPU count)

Tested (run the relevant ones):

• Code formatting: bash format.sh
• Any manual or new tests for this PR (please specify below)
• All smoke tests: pytest tests/test_smoke.py
• Relevant individual smoke tests: pytest tests/test_smoke.py::test_fill_in_the_name
• Backward compatibility tests: conda deactivate; bash -i tests/backward_compatibility_tests.sh

[cg505]

this has major issues after implementing parallelism control, most notably 1) lack of FIFO and 2) each pending job has a process alive with 70MB memory usage
superseded by #4485