feat(python): Expose infer_schema_length parameter on read_database

[alexander-beedie]

Closes #15059.

If not using an Arrow-aware driver or the "schema_overrides" parameter, and a column starts with > 100 null values, we need to expose the "infer_schema_length" parameter to allow for more generous dtype inference. (The other options are preferred, but we still need to make this parameter available if they cannot be used).

[stinodego]

Seems fair enough. Just for my understanding though: why does read_database need to do schema inference at all? Don't databases have a strict schema that we can use? I guess we cannot account for all possible third party data types so we look at the data instead?

[alexander-beedie]

I guess we cannot account for all possible third party data types so we look at the data instead?

The query result's cursor "description" property has a type_code entry, but it's an absolute free-for-all as to what you find in it. Could be a string name, an integer code known to the backend, a python-native class, a driver-specific enum, etc. Translating it often requires deep backend/driver-specific knowledge.

I do have a "TODO" to improve this on our side, but it's non-trivial (which I know well, because I have written code that does exactly this at work, but it's not simple to do the same for Polars, as I have an extensive dtype-translation architecture running to thousands of lines of code by itself in place there ;)

(Also there are backends like SQLite that don't populate the cursor type_code entry at all, so that will always require introspection - the result of a query may have no resemblance to the schema of the underlying tables, so knowing the table schemas typically won't help that much unless it's a "SELECT *").

[stinodego]

Thanks for the explanation!