Add local edge source/destination shuffle functions and clean up shuffle utility functions.

[seunghwak]

This PR pulls updates from #4934 and #4931. Better be reviewed/merged after these PRs.

This PR was initially created to add forest pruning code. But re-purposed to avoid creating a PR that adds/updates two features (shuffling & forest pruning, forest pruning requires edge source/destination shuffling functions). A separate forest pruning PR will be created pulling updates from this PR.

This PR

• Added utility functions to shuffle local edge source/destination vertices (already resides in the local GPU based on edge partitioning) to the owning GPUs (by vertex partitioning). This can be achieved just using a sub-communicator (major communicator or minor communicator involving only a subset of the GPUs). This is more efficient than using the existing shuffle_vertices which uses the global communicator.
• Renamed public shuffle utility functions. Several public shuffle functions have external in their function names (e.g. shuffle_external_vertices), but we are using ext or int abbreviations to refer external (unrenumbered) and internal (renumbered) vertices elsewhere (e.g. renumbering/unrenumbering functions). This PR renames external in those functions to ext for consistency.
• Splitted graph_functions.hpp to graph_functions.hpp and shuffle_functions.hpp and moved shuffle utility functions to the new shuffle_functions.hpp.
• We have been calling detail space functions in many places even when there exists a public function doing the same job. This PR updates the codebase to invoke public functions whenever possible.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.