test: error when sqlglot tells us a compilation is unsupported

[NickCrews]

This is a sanity check. If sqlglot tells us we are generating invalid SQL, then we should be handling it better.

I'm not sure how CI will like this. Potentially this will reveal a lot of bad assumptions we have been making.

My other motivation here is that if we merge this, then we can push more logic down into SQLglot. For example, if tobymao/sqlglot#6377 gets merged, then we can rely on sqlglot to tell us that column.first(include_nulls=False) is illegal to compile on postgres (only column.first(include_nulls=True) is possible to correctly compile. Currently, we have to do these sorts of checks in our individual backend compilers, for example see #11311

[cpcloud]

Wow, I am surprised that only 2 backends failed with this enabled!

[NickCrews]

@cpcloud if I fix those failing tests then does this seem like a good idea?

[NickCrews]

I think I fixed the flink error, that was easy.

For trino, I started tobymao/sqlglot#6616 to fix one of the trino errors. I still need to work on the other trino error.

[NickCrews]

Repro for the trino error for playing around:

import ibis
from ibis.backends.sql.compilers.trino import TrinoCompiler
import sqlglot as sg

data = {"a": [[2], [4]]}

t = ibis.memtable(data, schema=dict(a="!array<int8>"))

predicate = lambda x, i: x + (i - i) > 1
# predicate = partial(lambda x, y, i: x > y + (i * 0), y=1)
filtered = t.a.filter(predicate)

sge = TrinoCompiler().to_sqlglot(filtered)
sge.sql(
    dialect="trino",
    pretty=True,
    copy=False,
    unsupported_level=sg.ErrorLevel.RAISE,
)

Edit: sqlglot warns when compiling a sge.Struct on trino/presto if the type of any of the fields can't be inferred

We trigger this when compiling ops.ArrayFilter on trino, and in the temp struct we go through, we just pass a bare identifier expression, which sqlglot can't infer the type of.

So the solution was to pass a type hint to sqlglot the type of this struct field. Maybe not ideal sqlglot behavior, I feel like it should be valid to compile an untyped field with no warning, but we can work around it.

[NickCrews]

OK @cpcloud all tests pass now!

One possible thing to think about here: There may be cases that work just fine, but still raise sqlglot errors, that we aren't seeing in tests. If we all of a sudden make these error, then we might just be breaking some poor user for no real reason. So, perhaps we want to only enable this strictness check in tests? eg by looking at some environment variable that is set in conftest.py?