Default to using direct GPU communication with tMPI

Functionality changes:

• remove old dev flags and simplify setup to allowing either all GPU communication direct or not (rather than halo and PP-PME separately);
• make direct communication default for thread-MPI and allow disabling it using GMX_DISABLE_DIRECT_GPU_COMM;

Unchanged functionality but code or behaviro affected by the dev flags restructuring:

• keep direct GPU communication with CUDA-aware MPI off by default as this codepath still needs more testing, allow enabling it using GMX_ENABLE_DIRECT_GPU_COMM;
• avoid changes in what is tested in CI (disable the feature where it would now default to enabled);
• clean up removed dev flags in CI and docs;
• clean up log output, including references to the old dev flags;

Refs #3950 (closed)