polars array reading inconsistent dtypes

[TeunHuijben]

Fix Array(Float64) columns being truncated to ints on SQL reads

Reading an array column back from a SQLGraph can silently drop the fractional part of the values, when the leading rows happen to hold whole numbers:

graph = td.graph.SQLGraph(drivername="sqlite", database=":memory:")
graph.add_node_attr_key("pos", pl.Array(pl.Float64, 2))
graph.bulk_add_nodes([{"t": 0, "pos": [50, 50]},     # whole numbers
                      {"t": 1, "pos": [1.5, 1.5]}])  # fractional

graph.node_attrs(attr_keys=["pos"])["pos"].to_list()
# -> [[50.0, 50.0], [1.0, 1.0]]     the 1.5 became 1.0

No error, no warning, just wrong numbers. Swap the two rows and it works, graph.nodes[id]["pos"] is correct, and the in-memory backends are fine — only the bulk read on SQL truncates.

Array columns are stored as pickled blobs, so _polars_schema_override deliberately leaves them out (declaring Array(Float64, 2) there gives all-None, I tried). The loss happens a step later in unpickle_bytes_columns: once the blobs are unpickled, the column is rebuilt with pl.Series(values) and no dtype, so polars infers it from the leading rows — pl.Series([[50, 50], [1.5, 1.5]]) gives List(Int64) and [[50, 50], [1, 1]]. _cast_columns then casts the already-truncated values to Float64. We knew the declared dtype all along, we just never passed it.

Changes

• unpickle_bytes_columns takes an optional dtypes mapping and builds each column with its declared dtype, falling back to inference, then to object (for Mask).
• Added SQLGraph._declared_column_dtypes(table, pickled=...), which now backs both _polars_schema_override (natively stored columns) and the new _pickled_column_dtypes (blob columns). Wired into all four read sites.
• Side benefit: pl.Object columns build directly as object instead of being inferred into a struct and cast back, which drops a raise-per-read on mask columns.

Tests

test_array_attr_read_honors_declared_dtype, parametrized over all backends, covering Array(Float64, 2) and List(Float64) with a whole-numbered first row. Fails on SQLGraph only before the fix.