fix(connections): cancelled connect no longer aborts a later successful connection (#1358)

[datlechin]

Fixes #1358.

Cancelling a pending connection only marked the Swift Task cancelled; the blocking PQconnectdb kept running. When the host later became reachable, the cancelled connectToSession resumed and, with no cancellation guard, mutated activeSessions[connectionId], which by then belonged to the new successful attempt: either overwriting its driver or removing the live session.

Changes

• connectToSession checks Task.isCancelled after connect() and before the session write. This is the actual fix.
• A cancelled attempt disconnects only its own driver and never touches the connectionId-keyed session or SSH tunnel (which avoids tearing down the retry's tunnel).
• Genuine failures still clean up as before. Postgres connect() now uses the cancellable dispatch helper, which only short-circuits a cancel that lands before dispatch; it does not interrupt an in-flight blocking PQconnectdb, so the Task.isCancelled checks are what guarantee correctness.

Tests

CancelledConnectionCleanupTests: a cancelled attempt leaves the session intact; a genuine failure removes it and either clears currentSessionId or moves it to a remaining session. The race is timing-dependent, so these pin the extracted cleanup helper's behaviour rather than reproducing the race. All pass.

Known follow-up

Cancelling a tunnelled connection without retrying can leave the SSH tunnel up until a later attempt replaces it. Tracked separately.