fix(mysql): harden protocol decoders against malformed packets from hostile servers

[portlandhodl]

Found during a security review of server-controlled parsing paths (threat model: rogue or MITM MySQL/MariaDB server; handshake parsing is reachable pre-authentication).

Does your PR solve an issue?

No tracking issue — reported here directly. This extends the fix pattern of #4176 (truncated OK packet) to the rest of the protocol; #4451 documents the same review on the PostgreSQL side.

The problem

• recv_packet's continuation loop accumulated an unbounded number of 0xFFFFFF-sized parts: a hostile server streams endless 16 MiB frames and the client OOMs at near-zero bandwidth cost.
• The result-set column count (a length-encoded u64) fed HashMap::with_capacity / Vec::reserve directly; a ~10-byte packet claiming 2^40 columns aborts the process.
• Dozens of decode sites call panicking bytes::Buf getters / slice indexing / advance on untrusted buffers: truncated ERR packets, every field of HandshakeV10 (including the attacker-controlled auth-plugin-data length), length-encoded strings claiming more than the remaining bytes, TextRow field-count mismatch, BinaryRow (null bitmap, temporal length byte, fixed-width and lenenc value bounds — plus an unreachable!() actually reachable via a NULL-typed column), short EOF packets, and a 1-byte caching_sha2_password continue packet. The OkPacket decoder was hardened in fix(mysql): return error instead of panic on truncated OK packet #4176; the rest of the protocol was not.

The change

• Cumulative 1 GiB cap on packet reassembly (MySql's own max_allowed_packet hard ceiling); protocol error beyond it.
• Result sets claiming more than 4096 columns (the MySQL server hard limit) are protocol errors before any allocation.
• All listed sites check remaining length up front and return Error::Protocol describing expectation vs. reality; the unreachable!() is now an error.

Is this a breaking change?

No: every changed path converts a panic or abort into a returned Err, which callers already handle (the pool discards the connection). Newly-observable edges: aggregate packets over 1 GiB and result sets over 4096 columns are now errors rather than allocation attempts.

Regression tests

15 new unit tests, including a sweep asserting that every strict prefix of captured MySQL 8.0.18 and MariaDB 10.4.7 handshake packets now errors instead of panicking. Each test fails (panics) without the fix. cargo test -p sqlx-mysql --lib passes.