bench: add connection and query benchmarks for SQLite, Postgres, MySQL

[Map130]

Closes #158.

Adds criterion benchmarks covering all items from the issue checklist, for all three supported databases.

What's included

Each database gets a benches/<db>/connection.rs with five benchmarks:

Benchmark	Description
new_connection	Establish a fresh connection from scratch (connect only; graceful close is done outside the timed section)
pool_checkout	Acquire an idle connection from a pool (min=max=1, test_before_acquire(false))
ping	Ping an existing open connection
query_small_result	Fetch 1 row by PK from a populated table (TechEmpower single-query style)
query_large_result	Fetch 10 000 rows with TEXT and BLOB/BYTEA columns

How to run

SQLite (no external DB needed):

cargo bench --bench sqlite-connection --features sqlite,runtime-tokio

Postgres:

DATABASE_URL=postgres://... cargo bench --bench postgres-connection --features postgres,runtime-tokio

MySQL (mysql-rsa enables caching_sha2_password auth without TLS for MySQL 8+):

DATABASE_URL=mysql://... cargo bench --bench mysql-connection --features mysql-rsa,runtime-tokio

The existing tests/docker-compose.yml can be used to spin up Postgres and MySQL.

Notes on methodology

• pool_checkout uses test_before_acquire(false). With the pool default (true), every acquire() pings the connection before handing it out, which would make this benchmark measure ping + checkout rather than the checkout fast path. Disabling it isolates the pool's own overhead.
• new_connection times connect() only. Connections are closed gracefully (Terminate / COM_QUIT) outside the measured section via iter_custom, so teardown isn't included in the timing and we don't abandon thousands of sockets on the server over a run.
• Fetched results are passed through std::hint::black_box so the optimizer can't elide the decode/allocation work being measured.
• SQLite numbers aren't directly comparable to Postgres/MySQL. sqlite::memory: has no network round-trip, so its new_connection/ping/pool_checkout measure in-process setup rather than a real connection/RTT. The three DBs share the same schema (16-byte blob, 10 000 rows) but should be read per-database.