CORE: add global init status check

[wfaderhold21]

What

Extends an opt-in asymmetric-datatype check for rooted collectives to include a status check of ucc_coll_init ensuring all ranks either pass or fail.

Why ?

Previously a rank could fail ucc_coll_init and return an error while other ranks may succeed and hang waiting on the failed process.

[greptile-apps]

Greptile Summary

This PR extends the optional check_asymmetric_dt mechanism so that all ranks participating in gather/scatter collectives always call ucc_service_dt_check, regardless of whether their local ucc_coll_init succeeded. The local init status is encoded in two new int16_t slots (values[4,5]) in the allreduce buffer using the same min/max trick, allowing all ranks to agree on the global outcome rather than having failed ranks silently exit and leave others hanging.

• ucc_coll.c: Removes the early-exit after ucc_coll_init for gather/scatter types when check_asymmetric_dt is enabled, passing local_status and a possibly-NULL task to the revised ucc_service_dt_check signature.
• ucc_service_coll.c: Expands the allreduce payload from 4 to 6 int16_t values, adds a manual schedule-initialization path for the NULL-base_team case, and improves error-path task-completion handling.
• ucc_coll_utils.c: Fixes ucc_copy_asymmetric_buffer to use task->bargs.team->size (ucc_team_t *, always valid) instead of task->team->params.size (ucc_base_team_t *, now potentially NULL).

Confidence Score: 4/5

Safe to merge; the core hang-prevention logic is correct and force-completion paths are carefully guarded against double-completion.

The scheduling changes correctly encode init status in two extra int16_t slots and always route gather/scatter through ucc_service_dt_check, eliminating the targeted hang. The force-completion pattern correctly prevents ucc_schedule_completed_handler from triggering a second completion. The ucc_coll_utils.c fix correctly switches to task->bargs.team. The change touches several interacting paths in the collective scheduling mechanism, warranting careful review.

src/core/ucc_service_coll.c — the force-completion logic in ucc_dt_check_allreduce_post and ucc_dt_check_actual_wrapper_post interacts non-obviously with the UCC_EVENT_COMPLETED_SCHEDULE firing path.

Important Files Changed

Filename	Overview
src/core/ucc_coll.c	Removes early-exit guards around ucc_coll_init for gather/scatter types when check_asymmetric_dt is enabled; all ranks now reach ucc_service_dt_check regardless of local status. Scratch-buffer cleanup before the call is correct.
src/core/ucc_service_coll.c	Allreduce payload grows to 6 int16_t slots; manual schedule init for NULL-base_team case; force-completion paths set n_completed_tasks = n_tasks before calling ucc_task_complete on the schedule, correctly preventing double-complete from the subsequent UCC_EVENT_COMPLETED_SCHEDULE.
src/core/ucc_service_coll.h	Updated docstring and signature for ucc_service_dt_check; clearly documents the must-be-called-on-every-rank contract.
src/schedule/ucc_schedule.h	ucc_dt_check_state gains a 6th slot pair for init-status encoding and a new ar_status field for service-allreduce failures.
src/utils/ucc_coll_utils.c	Fixes ucc_copy_asymmetric_buffer to use task->bargs.team->size (always valid) instead of task->team->params.size (now potentially NULL).

_{Reviews (4): Last reviewed commit: "REVIEW: extend dt check to include init ..." | Re-trigger Greptile}