Add bulk_insert dialect hook with DuckDB CSV implementation

[martinv13]

Introduces DatabaseDialect.bulk_insert(conn, table, records) as the single insertion point for temp-table loading. The base implementation falls back to SQLAlchemy executemany (no behaviour change for PostgreSQL, MySQL, MSSQL). DuckDBDialect overrides it with a write-to-tempfile / read_csv approach that is significantly faster for large payloads:

• Records are serialised to a NamedTemporaryFile CSV (stdlib csv, no extra dependencies).
• read_csv is called with all_varchar=true; each column is then explicitly CAST to its target DuckDB type (BIGINT, DOUBLE, TIMESTAMPTZ, BOOLEAN, …) in the SELECT clause, avoiding auto_detect type mis-identification.
• LargeBinary (record-hash) columns are hex-encoded in the CSV and decoded with unhex() in SQL.
• SQLAlchemy Python-side scalar defaults (e.g. default=False on temp_exists) are materialised manually before writing the CSV, matching the behaviour of executemany.
• The temp file is deleted in a finally block even when an error occurs.

document.py: insert_into_temp_tables now calls
dialect.bulk_insert(conn, query.table, records) instead of conn.execute(query, records) directly.

Tests: new tests/test_bulk_insert.py covers base-class fallback, numeric types (incl. BigInteger/SmallInteger subclass ordering), boolean, datetime, binary, scalar defaults, and empty-records no-op.