Converge on gmxapi Context as the only way to manage running a simulation

There are currently three ways to manage running a simulation within the current source:

• by invoking a gmxapi::Workflow on a gmxapi::Context
• by calling gmx mdrun
• by using gmx::test::SimulationRunner in the mdrun integration tests

There's noteworthy duplication between the first and both of the subsequent two. Having to consider all three use cases complicates #3688 because responsibility for initializing MPI libraries is handled in three places. Likewise #3768 (closed) needs similar changes in three places.

In the long term, we should use only one way to run a simulation. Simplest and most generalizable is to re-use the gmxapi implementation for the other two. That prepares for re-using some of the infrastructure currently used for testing to provide capabilities to the user also via the API.

Tasks have been identified to achieve this:

• Discuss with nblib any efforts to re-use or further de-duplicate
• Make gmx::test::SimulationRunner use a gmxapi::Context to run mdrun
• Identify an approach that lets gmx::test::SimulationRunner call grompp
• Make gmx::test::SimulationRunner use that approach to call grompp
• Understand whether the apparent need to broadcast gmx command-line arguments before invoking gmx_mdrun can be resolved within the approach of gmxapi::Context
• Have gmx infrastructure implement gmx mdrun by creating a gmxapi::Context and transforming the command-line arguments to a gmxapi::Workflow
• Move sole responsibility for MPI library initialization to gmxapi::Context (see also #3688)
• Eliminate any superseded infrastructure, e.g. gmx::init()/finalize()