feat: Add Database#serialize and Database#deserialize methods

[Maaarcocr]

Add support for sqlite3_serialize() and sqlite3_deserialize() which
allow serializing an in-memory database to a binary string and
restoring it later.

These methods are useful for:

• Creating snapshots of in-memory databases
• Transferring database contents between connections
• Backing up and restoring database state

The methods are only available when SQLite is compiled without
SQLITE_OMIT_DESERIALIZE (requires SQLite >= 3.23.0).

[flavorjones]

@Maaarcocr The test suite failure seems legit. Let me know if you need assistance getting unblocked!

[Maaarcocr]

@Maaarcocr The test suite failure seems legit. Let me know if you need assistance getting unblocked!

sorry for the delay, getting to this only now. the issue I have is that I can't currently replicate. I run the test with no issues on my machine. could it be an OS issue? or a sqlite version issue?

[flavorjones]

@Maaarcocr It was a SQLite version issue. The basic CI job builds against Ubuntu 24.04's system SQLite 3.45.1, where select 1 succeeds after deserializing garbage, so test_deserialize_invalid_data failed. The vendored SQLite 3.54.0 raises instead, which is likely why it passed for you.

Older SQLite has a worse problem, too: before 3.54.0, sqlite3_deserialize() returns SQLITE_OK while a statement is unfinished, and the next step of that statement reads freed memory. SQLite 3.54.0 returns SQLITE_BUSY instead. So Database#deserialize is now defined only when the loaded SQLite is 3.54.0 or later, and that test now skips on the basic job.

The other changes I pushed on top of yours:

• I padded the deserialize buffer with 20 zeroed bytes, as the sqlite3_deserialize() docs advise. That also fixed deserialize(""), which raised NoMemoryError because sqlite3_malloc64(0) returns NULL.
• serialize now raises SQLite3::Exception for an unknown database name, instead of NoMemoryError: failed to allocate -1 bytes.
• The deserialize docs now say when it raises BusyException, that it detaches a file-backed database from its file, that a WAL-mode image can't be queried afterward, and that deserializing untrusted data is as risky as opening an untrusted database file.
• I added a CHANGELOG entry and fixed two -Wformat warnings.