[selmanozleyen]

Stream(group_by=()) now means one group holding every cell, with leaf (). sckitflow's unconditional schema currently adds a dummy "all" obs column to every AnnData only because group_by had to be non-empty; this lets it drop that.

• leaf_codes returns one leaf () for no columns; leaf categories are built with pd.Index(leaves, tupleize_cols=False) (an empty tuple is not a MultiIndex entry).
• materialize_node skips the sort when there is nothing to sort by.
• Tests: training across two files, an in-memory ungrouped link, and an eval pass over the one group.

Note: from_paths with only ungrouped streams reads no obs columns, which load_backed_adata treats as "read the full obs"; correct, but slow on very large stores.