New integration tests framework

[jngrad]

Fixes #3883

Description of changes:

• write more thorough RNG checks for thermostats
• skip slow/statistical tests in coverage and sanitizer builds
• fix broken ifdefs for Brownian Dynamics
• cleanup IBM code

[jngrad]

Here is the summary for the python testsuite, using the same CI runner when possible:

CI job	Old runtime	New runtime	speed-up
clang-sanitizer	2109.2s	988.1s	19m (53%)
no_rotation	647.7s	468.5s	3m (28%)
maxset	961.6s	616.0s	6m (36%)
default	970.5s	771.3s	3m (20%)
cuda11-maxset	1113.4s	774.9s	6m (30%)

The complete pipeline runtime went from 72 min to 38 min. The Clang job completes in 35 min instead of 50 min.

@RudolfWeeber your plan in #3883 (comment) was really good, however I had to deviate from it slightly:

• virtual_sites_tracers is needed to get code coverage for IBM particle coupling to the fluid (not sure why though, the LB actor is removed before the IBM bonds are created)
• Brownian/NpT were moved fully, but new files with minimum checks were introduced for both integrators, otherwise their code coverage would be 0%

Other remarks:

• make check_python still works the same way as before
• make check_python_skip_long to run the subset of fast tests
• the comment lines in the thermostat tests about RNG seed and counter state were obtained using cout.diff

[jngrad]

The slowest tests are now:

• stokesian_dynamics: 1 min 10 sec in sanitizer, 40 sec in coverage, 20-30s in others
  • now skipped in sanitizer/coverage builds
  • new test added to check the SD interface
• elc_vs_analytic: 1 min 20 sec in sanitizer/coverage
  • removing it from coverage builds would require writing a small test without constant potential to preserve code coverage
• reaction ensemble tests:
  • constant_pH: 1 min 30 sec in sanitizer, 45 sec in coverage
  • reaction_ensemble: 30-40 sec in sanitizer/coverage, 5 sec in others
  • widom_insertion: 30-40 sec in sanitizer/coverage, 5 sec in others
  • wang_landau_reaction_ensemble: already skipped in sanitizer/coverage
  • long-term plan would be to skip these tests in sanitizer/coverage and write much smaller tests to check the RE interface (error messages, output files, and run a few reaction steps), like I did here for the thermostats and integrators

[RudolfWeeber]

I thought it doesn't matter, where in the system the gap is? @reinaual
Did this change influence the test outcome?

[jngrad]

This change did not alter the numerical result. I think the gap is always at the higher value of the z-axis:

espresso/src/core/electrostatics_magnetostatics/elc.hpp

Lines 92 to 93 in 127ad4e

	/** Up to where particles can be found. */
	double h;

espresso/src/core/electrostatics_magnetostatics/elc.cpp

Line 1196 in 127ad4e

elc_params.h = box_geo.length()[2] - gap_size;

If particles enter the gap, they are misclassified as particles from the dielectric contrast when calculating the forces and energies. The documentation warns against using ELC without walls to block particles from entering the gap:

espresso/doc/sphinx/electrostatics.rst

Lines 254 to 259 in 127ad4e

	* The method relies on a slab of the simulation box perpendicular to the
	z-direction not to contain particles. The size in z-direction of this slab
	is controlled by the ``gap_size`` parameter. The user has to ensure that
	no particles enter this region by means of constraints or by fixing the
	particles' z-coordinate. When there is no empty slab of the specified size,
	the method will silently produce wrong results.

[reinaual]

In theory it should not matter, to my knowledge in the implementation it does. The lower interface is assumed at z=0 and the upper one at z=h, which is relevant for the error-estimation and the placement of the image charges.
To my understanding the testcase should be correct for every given prefactor, however setting it to 1e-100 allows to explicitly test the potential difference contribution but hides the misplaced particles in this case.