[Bug]: -2**63 // -1 inconsistent across backends, SQLFrame silently returns a float #4071

Describe the bug

Follow-up from #3994 as suggested by @FBruzzesi. Related to #2748.

The value -2**63 // -1 overflows int64, which backends handle in different ways. Polars, pandas and Dask wrap around to -2**63. pyarrow and DuckDB raise an overflow error, and ibis raises a conversion error. Finally, SQLFrame returns 9.223372036854776e+18 as a float with no error.

I think the SQLFrame one is the result I'd worry about most, since it's silently wrong.

Steps or code to reproduce the bug

Simply running a file like this is enough to see the bug:

import narwhals as nw
import polars as pl
from sqlframe.duckdb import DuckDBSession

expr = nw.col("a") // nw.col("b")
print(nw.from_native(pl.DataFrame({"a": [-(2**63)], "b": [-1]})).select(r=expr)["r"].to_list())

df = DuckDBSession().createDataFrame([(-(2**63), -1)], ["a", "b"])
print(nw.from_native(df).select(r=expr).collect()["r"].to_list())

Expected results

I'm not sure which behaviour narwhals is expected to adhere to here. Wrapping would match Polars, and raising is what pyarrow and DuckDB do. But either way SQLFrame shouldn't be returning a float.

Actual results

[-9223372036854775808]
[9.223372036854776e+18]

Across all backends (narwhals main 455269e):

Please run narwhals.show_versions() and enter the output below.

Python dependencies:
     narwhals: 2.26.0
        numpy: 2.5.3
       pandas: 3.0.6
        modin: 
         cudf: 
      pyarrow: 25.0.1
      pyspark: 
       polars: 1.44.2
         dask: 2026.8.0
       duckdb: 1.5.5
         ibis: 12.0.0
     sqlframe: 4.4.0

Relevant log output