fix: any and all functions should not have default frame or order_by

[ssabdb]

Description of changes

The any and all operators should not have a default ordering or set a restricted 'current row' frame by default, because they are not operators which by default should be ordered.

I noticed a bug in bigquery where filtering e.g.
table.filter((.a == "hello").any().over(group_by=(.x)))

produces an incorrect window clause, including
(PARTITION BY t0.x order by null asc ROWS between unbounded preceding and current row)

which leads to inconsistent and incorrect results.

The expected behaviour is to produce no frame or order by, e.g.

(PARTITION BY t0.x ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING)

This is because ordering by null and the current row results in a undefined order. I suspect this will probably apply to multiple backends.

The fix involves modifying exclude_unsupported_window_frame_from_ops in rewrites.py to exclude Any and All.

Closes #11931

[ssabdb]

Since the change is to the base sql compiler rewrites.py, should the test be moved to ibis/backends/sql/tests/test_compiler.py?

[NickCrews]

Since the change is to the base sql compiler rewrites.py, should the test be moved to ibis/backends/sql/tests/test_compiler.py?

Almost, I think that file is just for testing the base Sql compiler, when ideally we would actually want to test this for each backend. So, it should go in ibis/backends/tests/test_sql.py, eg toss this at the bottom

@pytest.mark.parametrize("backend_name", _get_backends_to_test())
@pytest.mark.notimpl(["polars"], raises=ValueError, reason="not a SQL backend")
def test_window_filter_any(backend_name, snapshot):
    t = ibis.table([("a", "string"), ("x", "int64")], name="t")
    expr = t.filter((_.a == "hello").any().over(group_by=_.x))
    sql = ibis.to_sql(expr, dialect=backend_name)
    snapshot.assert_match(sql, "out.sql")

Second, can you actually explain what the exclude_unsupported_window_frame_from_ops rewrite rule DOES? I am decently confident in this change because the tests pass, but I don't have a good grip on what the other side-effects of this change would be. If you understand, it would be great if you could add a comment to the rewrite func

[NickCrews]

See my above comment, move the test to check all backends, and ideally explain the rewrite rule

[ssabdb]

Many thanks for the comments.

I'm also confused by what some of these rewrites functions are trying to do with window functions and on moving the tests noticed that some backends, e.g. mysql were still creating a window function with an order by null.

I think exclude_unsupported_window_frame_from_ops is designed to exclude window functions from navigation functions, e.g. lag, lead where there is no meaning to a frame. Any and All operators were erroneously included.

The function exclude_unsupported_window_frame_from_rank also silently injects a null order by if it's missing for a rank and the rewrite_empty_order_by_window always injects a null order for all window functions for certain backends if they are missing, e.g. mysql if it makes sense or not for the operator being used.

My conclusion is that these functions need a small amount of rationalization, probably to separate one to zero index transformation - there seem to be several of these - from forcing default order_by which feels like it should be consistent across all backends?

I've pushed some (not ready for merging) changes which tries to work on the rewrites only. I suspect it will break a few things.