fix: worker restart collection mismatch in loadscope/loadgroup scheduler

[C1-BA-B1-F3]

Description

This PR fixes two related issues when a worker crashes and is restarted with --dist loadscope or --dist loadgroup:

Issue #1189: KeyError when replacement worker's collection doesn't match

When a worker crashes, its entry in registered_collections was not cleaned up. This caused:

1. collection_is_completed to remain True even though the crashed node was gone
2. Replacement workers whose collections didn't match the original would have their collections silently dropped (the return statement in add_node_collection())
3. When schedule() tried to assign work to the replacement worker, it would raise KeyError because the worker wasn't in registered_collections

Issue #1323: Hang when completed work units are requeued

When a worker crashed, ALL work units (including completed ones) were added back to the workqueue via self.workqueue.update(workload). When these completed work units were later assigned to a new worker:

1. The filtering in _assign_work_unit() would produce an empty nodeids_indexes list (all items marked as completed)
2. node.send_runtest_some([]) would be called with an empty list
3. The worker would hang waiting for work that never arrives

Changes

src/xdist/scheduler/loadscope.py

1. remove_node(): Now removes the crashed node from registered_collections using self.registered_collections.pop(node, None). This allows replacement workers to properly register their collections.

2. remove_node(): Only requeues work units that have pending (uncompleted) items. Completed work units are filtered out before adding to the workqueue.

3. add_node_collection(): Always registers the collection even if it doesn't match the existing collection. The mismatch will be properly reported by _check_nodes_have_same_collection() when schedule() is called.

Testing

All existing tests pass (220 passed, 6 skipped, 10 xfailed).

Related Issues

Fixes #1189
Fixes #1323