feat(snowflake): support VECTOR type via Array(length=N)

[daveauerbach]

Summary

Snowflake's VECTOR(<element>, <dimension>) type stores fixed-length numeric vectors used by AI/ML workloads. The Snowflake Python connector already deserializes VECTOR columns to native pyarrow fixed_size_list<element>[length], but ibis had no VECTOR -> ibis dtype mapping, so the column came back typed as utf8 (via the Unknown -> string fallback) and SnowflakePyArrowData.convert_column crashed downstream with:

pyarrow.lib.ArrowNotImplementedError: Unsupported cast from
fixed_size_list<element: float not null>[N] to utf8 using function cast_string

This PR adds three pieces:

1. SnowflakeType._from_sqlglot_VECTOR parses Snowflake VECTOR into dt.Array(<element>, length=<dimension>). Per the Snowflake docs, element types are 32-bit (INT -> 32-bit signed integer, FLOAT -> 32-bit single-precision floating-point). sqlglot normalizes the VECTOR-element FLOAT to Type.DOUBLE in its DataType AST, so we narrow to Float32/Int32 rather than falling back to the default scalar mappings (which would give Float64/Int64). VECTOR elements are also not nullable per Snowflake's storage model, hence nullable=False on the inner dtype.

2. SnowflakePyArrowData.convert_column / convert_scalar now pass fixed-length array columns through unchanged. Without this guard, the JSON-extension wrapping path tries to cast the fixed_size_list<...> data to utf8 (the storage type of PYARROW_JSON_TYPE) and raises ArrowNotImplementedError.

3. SnowflakeType._from_ibis_Array round-trips fixed-length numeric Arrays back to VECTOR(<element>, <dimension>). Other fixed-length element types (e.g. strings, structs) and variable-length arrays continue to compile to ARRAY (the existing JSON-backed variant-array).

Test plan

• ibis/backends/snowflake/tests/test_datatypes.py::test_parse adds three parametrized VECTOR cases (VECTOR(FLOAT, 4), VECTOR(FLOAT, 512), VECTOR(INT, 8)).
• test_extract_type_from_table_query adds two parametrized VECTOR cases that round-trip through a live Snowflake CREATE TEMP TABLE.
• New test_vector_column_pyarrow_passthrough_for_fixed_size_arrays exercises the converter directly with a synthetic pyarrow fixed_size_list<float32>[4] column -- no Snowflake connection needed -- and asserts both pass-through identity and value preservation.

Locally:

$ pytest ibis/backends/snowflake/tests/test_datatypes.py -v -k "test_parse or test_vector_column_pyarrow"
======================== 18 passed in 0.05s =========================

(The test_extract_type_from_table_query parametrized cases are gated on a live Snowflake connection and will run in CI.)

Notes

• Same pattern shape as the previous Snowflake compiler PR I shipped a few months ago, #11940 (which added visit_HexDigest for SHA2/MD5).
• Discovered while exercising a real PROD Snowflake table with a VECTOR(FLOAT, 512) column (an embedding store) through ibis.snowflake.to_polars().

Made with Cursor

Made with Cursor