cancelled `begin()` leaves pooled Postgres connection with an open transaction

[imabdulbasit]

Description

PgTransactionManager::begin queues BEGIN, then awaits wait_until_ready() before incrementing transaction_depth:

let rollback = Rollback::new(conn);
rollback.conn.queue_simple_query(&statement)?;
rollback.conn.wait_until_ready().await?;        // cancellation window
if !rollback.conn.in_transaction() {
    return Err(Error::BeginFailed);
}
rollback.conn.inner.transaction_depth += 1;     // never reached on cancel
rollback.defuse();

If the future is dropped at that await (task abort, select!, tokio::time::timeout), the BEGIN is already in the connection's write buffer or on the wire, but depth is still 0, so the Rollback guard's start_rollback no-ops (if conn.inner.transaction_depth > 0). No ROLLBACK is queued and the connection returns to the pool with an open server-side transaction.

In 0.8.3 the increment happened before the await, so a cancelled begin left depth = 1 and the guard queued the compensating ROLLBACK. The reorder came in with #3765 (merged for 0.8.4) and is still present in 0.9.0.

Reproduction steps

1. Create a PgPool with max_connections(1) so the same connection must be reused.
2. Start pool.begin() but drop the future before it completes.
3. Use the pool again: the next pool.begin() on that connection elicits WARNING: there is already a transaction in progress
4. Run the same sequence on 0.8.3: no warning, because the drop guard queues the compensating ROLLBACK.

SQLx version

0.8.6 (regression present in 0.8.4 through 0.9.0; 0.8.3 not affected)

Enabled SQLx features

postgres, runtime-tokio, tls-native-tls, chrono, bit-vec

Database server and version

PostgreSQL (reproduces against the official postgres Docker image)

Operating system

Linux x86_64; also reproduces on macOS

Rust version

rustc 1.97.1