Feat: allow user-supplied queries to generate unit test data

[georgesittas]

Fixes #2542, TL;DR this PR adds a new format for specifying unit test fixture/expected data: SQL. :-)

• Discuss identifier normalization decisions. Initial approach was to execute user-supplied queries without normalizing or quoting them, but this led to issues in the create_view method because the fully-qualified view name in the CREATE VIEW <name> AS ... statement won't be quoted if we set quote_identifiers=False, which may cause problems.

@erindru I ran your example against Trino and it seemed to work fine. I feel like the flexibility this feature provides will allow users to bypass conversion issues in unit tests pretty easily. Thanks for the idea!

EDIT: will make a followup PR to update the unit test documentation.

[tobymao]

this is confusing, but i think it makes sense, we need to normalize by the model, that's always true or else references will be wrong.

the parsing should be supplied by the actual test engine so there are no nuanences with transpilation

[georgesittas]

both points are exactly what I had in mind 👍

[erindru]

Allowing users to supply arbitrary SQL to produce a resultset for tests is a nice escape hatch imo.

Of course in an ideal world SQLGlot would be able to generate the correct query 100% of the time, but the variations in SQL syntax and obscure data types between engines are essentially infinite.

I think allowing users to un-block themselves by supplying the query manually when they encounter an edge case will go a long way to reducing friction