Improve multinode tutorial

[95-martin-orion]

A working version of the multinode tutorial has been submitted to the repo, but we should continue to improve upon it. Below is a short (incomplete) list of changes that would make the tutorial smoother for the user.

• Pre-built images for the compute nodes. This would considerably reduce bringup times, which currently consume 10-15 minutes of the tutorial time.
• GPU support. This essentially requires #430, as setting this up without a PyPI package requires a bunch of extra effort on the user side to build (and possibly distribute) GPU images.
• Seed non-collision guarantees. The current tutorial relies on the large seed space to prevent collisions between seeds selected by nodes. This gives a low, but crucially non-zero risk of collision. To resolve this, we need to remove a redundant randomizing layer in the simulator code and update the tutorial to cover explicit seed distribution.

[95-martin-orion]

Related: #454