Add vessel-graph connected-component labeling and node annotation rule

[Copilot]

Vessel nodes/edges are now generated, but nodes lacked graph connectivity context. This change adds a workflow step that computes connected components from the vessel graph and annotates each node with a component label ranked from largest to smallest.

• Workflow graph updates

  • Split node generation into two stages:
    • vessel_graph_to_nodes_edges now emits nodes_raw.parquet + edges.parquet
    • new vessel_graph_connected_components consumes nodes_raw.parquet + edges.parquet and writes final nodes.parquet
  • Preserves existing final artifact naming (nodes.parquet) while introducing an explicit intermediate node table.

• Connected-components annotation

  • Added spimquant/workflow/scripts/annotate_vessel_graph_connected_components.py.
  • Computes connected components from src_node_id/dst_node_id and adds component_label to nodes.
  • Labels are assigned by descending component size (1 = largest), with deterministic tie-break by minimum node_id.

• Compatibility + robustness

  • Updated convert_vessel_graph_to_nodes_edges.py to support both output contracts (nodes_raw_parquet preferred, nodes_parquet fallback) for backward compatibility.
  • Added guards for malformed edge tables (missing endpoint columns / unknown node references).

# Component ranking policy in annotate_vessel_graph_connected_components.py
ranked_components = sorted(
    components_by_root.values(),
    key=lambda comp_node_ids: (-len(comp_node_ids), min(comp_node_ids)),
)
for label, component_node_ids in enumerate(ranked_components, start=1):
    for node_id in component_node_ids:
        component_label_by_node_id[node_id] = label

[akhanf]

This ends up being too memory inefficient - ideally should move to using cugraph to do this, making use of a GPU -- however, currently dask-cuda pins to an older version of dask than zarrnii (zarrnii is pinned to newer version to support sharding).. so may need to 1) implement the vessel graph processing in a distinct conda env for now, 2) fork dask-cuda and remove the pin, or 3) wait for next dask-cuda release where the pin will move.