Make Steepest Descent a regular integrator

[jngrad]

Description of changes:

• Remove the deprecated minimize_energy module
  • use system.integrator.set_steepest_descent() directly
• Simplify the MPI callback mechanism for Steepest Descent
  • broadcast struct parameters only once during initialization, instead of once per integration
  • call the integrate() MPI callback directly instead of delegating that call to the steepest_descent() MPI callback
• Document how to set up a custom convergence criterion

[review-notebook-app]

Check out this pull request on 

Review Jupyter notebook visual diffs & provide feedback on notebooks.

---

Powered by ReviewNB

[jngrad]

This PR depends on #3880 and currently includes it. Here is the relevant diff subset.
@christophlohrmann could you please have a look at the samples and tutorials?

samples/chamber_game.py
samples/drude_bmimpf6.py
samples/visualization_bonded.py
samples/visualization_cellsystem.py
samples/visualization_charged.py
samples/visualization_elc.py
samples/visualization_npt.py
doc/tutorials/02-charged_system/02-charged_system-1.ipynb
doc/tutorials/02-charged_system/02-charged_system-2.ipynb
doc/tutorials/02-charged_system/scripts/nacl.py
doc/tutorials/02-charged_system/scripts/nacl_units_confined.py
doc/tutorials/02-charged_system/scripts/nacl_units_confined_vis.py
doc/tutorials/04-lattice_boltzmann/04-lattice_boltzmann_part3.ipynb
doc/tutorials/04-lattice_boltzmann/scripts/04-lattice_boltzmann_part3_solution.py

[christophlohrmann]

From manual check: Visualizers look good, systems do not explode

[KaiSzuttor]

LGTM except for the long testcase that needs to be shortened somehow.

[jngrad]

There is a silent regression somewhere. I get random failures in the ROCm image for tests dawaanr-and-dds-gpu and dawaanr-and-dds-gpu that look like UB (logfiles 1, 2, 3), but I can't reproduce them on lama nor in the ROCm docker image on lama. They are only reproducible in the docker image on coyote10, but not every time, and in GDB the backtrace is sometimes squashed. The few times I got a usable backtrace in GDB, it was about incorrect particle ids and double free of the cellsystem particle list (dawaanr-and-dds-gpu traces 1.log, 2.log).

The regression is independent of 022ee4e and independent of another silent regression that was fixed by 112bd65.

[jngrad]

Even more failures after the bugfix in 112bd65 (logfile):

	 32 - dawaanr-and-dds-gpu (Failed)
	 33 - dawaanr-and-bh-gpu (Failed)
	 34 - dds-and-bh-gpu (Failed)

This is not a regression, the bug is already present on the python branch. Tested on fe967ea with the ROCm image on coyote11: dawaanr-and-dds-gpu, dawaanr-and-bh-gp and dds-and-bh-gpu all fail randomly. It's also failing in other PRs.

[KaiSzuttor]

after my comments have been addressed, this can be merged