Fix detection of parquet filter pushdown

[phofl]

we don't want to use are_co_aligned here, it's way too permissive.

    # Don't push down if replace is in there
    x = df[df.replace(5, 50).a == 5]["b"]
    y = x.optimize(fuse=False)
    assert isinstance(y.expr, Filter)
    assert len(y.compute()) == 0

This shouldn't move the filters inside of read_parquet

Another lessons learned for me personally is that we shouldn't use it in hot-paths in the current implementation

[fjetter]

is this still needed with #909 (comment) ?

[phofl]

PR title is bad (sorry), this is actually more of a bug fix that brings a performance improvement as well. Currently we are too eagerly pushing down filters into the read_parquet call (see the example in the top post). are_co_aligned will allow all kinds of operations between read_parquet and the actual predicate, which will produce incorrect results

[fjetter]

good catch