test(contrib/drivers/sqlite): expand test coverage to the MySQL baseline + 2 driver fixes - #4892

ref #4689

Summary

Brings the SQLite driver up to the MySQL baseline. Test functions go from 161 to 561:

• the remaining feature suites ported from MySQL, every function keeping the name of its MySQL counterpart;
• sqlite_z_unit_feature_sqlite_test.go, 35 functions for behaviour that is SQLite's own: storage classes and typeof, WITHOUT ROWID, JSON1, window functions, CTEs, RETURNING, ON CONFLICT, INSERT OR IGNORE/REPLACE, LIMIT -1, last_insert_rowid(), EXPLAIN QUERY PLAN, date/time functions, RANDOM() ordering, the extra DSN pragmas.

Nothing under database/ is touched.

Four suites are not ported, on purpose:

Test_Model_Schema1/2 are omitted for the same reason as master_slave.

Where SQLite defines behaviour differently, the expectations follow SQLite, verified by running rather than reasoned about: one isolation level, database is locked in place of deadlocks, storage classes in place of declared types, JSON1 functions in place of a JSON column type, busy_timeout + WAL for the multi-connection scenarios.

Bugs found

Porting surfaced five. Two are fixed here because the tests cannot pass without them and the fix is confined to the SQLite drivers; the others are tracked separately.

1. Union / UnionAll never worked on SQLite. The core builds (SELECT ...) UNION (SELECT ...); SQLite forbids parentheses around compound operands. Not one existing test touched Union, so this was never noticed. Fixed in DoFilter of both sqlite and sqlitecgo: each operand becomes SELECT * FROM (...), which keeps a per-operand ORDER BY/LIMIT legal too. Parentheses are matched, not pattern-guessed; string literals are skipped.

2. WhereExists / WhereNotExists fail on SQLite. Wheref("EXISTS (?)", subQuery) expands to EXISTS ((SELECT ...)), which SQLite rejects. Fixed in the same DoFilter: the doubled pair is reduced to one.

3. CacheOption.Force never caches an empty result — #4891. All() returns a nil Result for no rows, which skips the Force branch in saveSelectResultToCache. Core logic, reproduced identically on MySQL. Every driver carries an empty Test_Model_Cache_Force for this reason, with a note that misstates why; this one is the same stub but says why. The scenario gets asserted, in every driver, with the fix for #4891.

4. A column holding mixed storage classes is typed from its first row — #4895. Once the first row is an INTEGER, later REAL / TEXT / BLOB rows of the same column are coerced to int64 and read back as 1 / 0. Core logic (the ScanType fast path in columnValueToLocalValue), and specific to the modernc driver, which reports the first row's Go type as the column's scan type; sqlitecgo is not affected. Test_SQLite_Type_StorageClass_MixedRows is kept as a stub pointing to #4895 and gets its assertions with that fix; the per-row single-value reads are asserted in Test_SQLite_Type_StorageClass_Typeof.

5. A POINT column reads back as 0 — #4842, fixed by #4882 which has since merged; Test_DataType_Geometry_Point passes on the rebased branch.

Verification

Full suite on Windows 11 with an isolated temp dir, -count=1, on d924ae001:

561 functions: 561 pass, 0 fail

The DoFilter rewrite has its own unit tests (sqlite_z_unit_do_filter_test.go) covering the shapes the core emits, nested sub-queries, and parentheses inside string literals. The existing sqlitecgo suite passes with the shared rewrite.