Provisioner cleanup

[allenwang28]

This PR implements a few cleanups after #285

Changes:

• Introduces a utility for just get_remote_info() - given a host mesh, spawn a proc and a setup actor, get the host port and address of the hostmesh, and kill the proc. This happens regardless of the underlying launcher. Slight refactoring in the bootstrap
  • Launchers no longer return this remote info
• Monarch now complains (with AssertionError) if you try and create an empty workspace. This doesn't really apply for SLURM, so we just create a tempfile and pass it to the workspace
• Gets rid of controller.proc_mesh.get_proc_mesh etc. and redirects everything to just use provisioner.get_proc_mesh
• Add host mesh <> proc tracking. I basically thought that I could do multiple spawn_procs(...) but I see that the prior procs die still. The intention was to enable the colocation of the vLLM controller and worker. We do this tracking because Monarch doesn't yet let you get the host_mesh from the proc_mesh (but it will eventually)
• Also adds the capability to provide a host mesh to get_proc_mesh - this way you can colocate things

[felipemello1]

should it silently fail?

[Ritesh1905]

For now, yes. (However some sort of logging would be idea.) until we figure out what would be a right way to segregate meta-only code blocks. These dependencies are meta internal and are installed via an internal fbpkg monarch build. The env setup for mast requires a separate installation script.

[allenwang28]

I've just added in a block that says "if imports failed and you're trying to use MAST, print an error messaging say imports failed and that you should check your build was correct"

[felipemello1]

can we use the actual class, e.g. ProcMesh?

also add a typehint?

[Ritesh1905]

Do you need these both? what's the difference between these two?

[allenwang28]

great catch! Removed the DELIVER which was a typo