fix(sqlite): stop multi-statement queries copying their tail on every prepare

[andeki92]

Does your PR solve an issue?

fixes #4415

The query buffer now ends in a NUL terminator that every tail keeps, so sqlite3_prepare_v3 parses in place instead of copying the rest of the query for each statement. Executing N statements goes from quadratic to linear: 80,000 statements in one raw_sql call take 201 ms instead of 4.78 s.

• prepare() now makes one sqlite3_prepare_v3 call per invocation. prepare_next already calls it again when no statement comes back, and the one-call shape lets the regression test see every tail handed to SQLite.
• A query containing a NUL byte used to hang the worker thread. VirtualStatement::new now rejects it with Error::InvalidArgument, and prepare() returns an error instead of looping when SQLite consumes nothing.

Tests:

• every_tail_handed_to_sqlite_ends_in_nul (unit): the regression test. It fails before the fix.
• prepare_errors_when_sqlite_consumes_nothing (unit) and it_rejects_a_query_containing_a_nul_byte: both hang before the fix.
• it_executes_statements_followed_by_whitespace_or_a_comment and it_executes_a_query_that_is_only_a_comment: they pass before and after, and cover the new loop-exit check.

Run locally on the pinned 1.94.1 toolchain: cargo fmt --check, ./x.py --clippy (20/20, no warnings), and cargo test -p sqlx-sqlite --all-features. The full ./x.py matrix passes except for 29 MariaDB targets that fail the same way on main: the client_ssl URL in x.py is missing the leading . in ssl-key=%2Ftests…, and the 10.6 and 10.11 images don't enable TLS while x.py asks for ssl-mode=required. I ran the Postgres/MySQL/MariaDB targets in a Linux container, because macOS native-tls can't do a TLS handshake with the Ed25519 test certificates.

Is this a breaking change?

No. Any query that returned before returns the same results now. The only behaviour change is for queries containing a NUL byte, which used to hang forever and now return an error.

---

Worked on with the help of Claude Opus 5.5.

[andeki92]

The MySQL 8 failures look unrelated to this change: every connection in --test any gets EOF before the server greeting, in the step that drops ssl-mode=disabled. sqlx-sqlite isn't compiled in these jobs, and the same job failed identically on #4413 (run 35537003373), which changes nothing the MySQL job builds. Could a maintainer re-run the failed jobs?