ci: add python wheel smoke test

[leeloveloli]

Summary

• add a minimal vexdb_lite Python wrapper package that opens DuckDB connections with VEX loaded by default
• add a Python wheel smoke test covering extension load, distance functions, GRAPH_INDEX creation, filtered ANN query, and VEX_INDEX_SCAN planning
• update the wheel workflow to inject the current vexdb_duckdb + common layout into the DuckDB sdist and target DuckDB 1.5.2
• document the new Python wheel entry point

Tests

• python -m py_compile packaging\vexdb_lite\__init__.py packaging\smoke_test.py
• python -c "import yaml; yaml.safe_load(open('.github/workflows/build-wheels.yml', encoding='utf-8')); print('workflow yaml ok')"
• git diff --check

Note: I did not run the full wheel build locally because compiling the DuckDB Python wheel is heavy; the PR focuses on fixing the packaging files and smoke-test path that the existing workflow already references.

[vibeinging]

发现一个阻断问题，建议先改完再合并。

packaging/smoke_test.py 里新增的 smoke test 创建了：

CREATE INDEX idx_items_vec ON items USING GRAPH_INDEX (vec, category) WITH (m = 8)

但这个 PR 的目标分支 dev 中，vexdb_duckdb/index/graph_index.cpp 明确拒绝多列/filtered GRAPH_INDEX：当 input.unbound_expressions.size() > 1 时会抛出 GRAPH_INDEX currently supports only a single FLOAT[N] column; multi-column (hybrid/filtered) form is disabled in this release。

这会导致新加的 wheel smoke test 在 CREATE INDEX 阶段直接失败，还没走到后面的 ANN 查询和 VEX_INDEX_SCAN 校验。

建议把 smoke test 改成当前支持的 GRAPH_INDEX (vec)。如果想覆盖 WHERE category = ...，可以先验证单列索引 + 后置过滤路径；如果当前版本不能保证这个行为，就先覆盖 unfiltered ANN，把 filtered ANN 单独留到 multi-column/hybrid support 恢复后再测。

补充确认：VEX_INDEX_SCAN 和 vex_brute_force_threshold 在这个 PR 的 dev base 上都是正确命名，不是问题。