Cleanup the "decomposition" directory

TL;DR The icon4py.model.common.decomposition package has outgrown its name: it mixes grid decomposition, MPI runtime, halo exchanges, global reductions, and single-node fallbacks. We should reorganize it around the broader concept of distributed computation.

Problem / motivation

model/common/src/icon4py/model/common/decomposition currently contains several distinct concerns that are only loosely related to “decomposition” in the grid sense:

• Grid decomposition algorithms (MetisDecomposer, SingleNodeDecomposer).
• Halo geometry construction (IconLikeHaloConstructor, NoHalos).
• Distributed exchange / halo exchange runtime (ExchangeRuntime, GHexMultiNodeExchange, SingleNodeExchange).
• Global reductions (Reductions, GlobalReductions, SingleNodeReductions).
• Process topology and MPI communicator handling (ProcessProperties, MPICommProcessProperties, init_mpi).
• Run-type dispatching and single-node vs multi-node abstractions (RunType, SingleNodeRun, MultiNodeRun).

The directory name and file names (e.g. mpi_decomposition.py, definitions.py) do not reflect this breadth, making it hard to discover where a given concept lives.

Ideas to explore

• Rename the package and files to reflect “distributed computation” rather than only “decomposition”.
• Split definitions.py and mpi_decomposition.py into smaller modules grouped by concern: decomposition, halo geometry, distributed exchange, reductions, process/runtime setup.
• Separate MPI-specific code from generic protocols and single-node fallbacks more clearly.
• Reconsider whether grid decomposition belongs under the same package as MPI runtime utilities.
• Clarify naming around DecompositionInfo — it carries halo and ownership information, not just decomposition output.
• Revisit ExchangeRuntime as a generic distributed communication abstraction; evaluate if halo exchange and reductions should share a runtime surface.
• Consolidate single-node vs multi-node dispatch patterns (get_runtype, create_exchange, create_reduction, get_process_properties) in one place.

Open questions / conflicts

• What is the right top-level name? distributed? parallel? distributed_runtime?
• Should halo construction live with grid utilities or with distributed communication?
• How much of this overlaps with planned component / model-state cleanup? See Revive components.