Retry queries only before sending

[ruslandoga]

A completed INSERT could be silently repeated when its successful HTTP response was lost. Ch now returns connection errors without replaying any query once sending starts. A lost INSERT acknowledgement leaves exactly one stored row.

Before sending a new request, check open? and probe idle connections with recv(conn, 0, 0). Failures request a reconnect and retry through DBConnection, which retains ownership of the connection lifecycle. A zero-timeout result continues on the same healthy connection. Stream continuations bypass the probe.

Require Mint ~> 1.11 and lock 1.11.0, which keeps healthy connections open when the probe times out. Ordinary, Enumerable, and Collectable INSERT failures cannot replay a write or retry a chunk against an obsolete request reference. Cursor declaration failures return DBConnection's supported non-retry disconnect result. Queries using an existing DBConnection.run/3 connection return errors without automatically switching connections.

Validation:

• Regression coverage for committed INSERTs with dropped acknowledgements, stale-idle recovery before an INSERT, stream continuation failures, and read response timeouts or closes without replay.
• Full CI matrix passed, including Elixir 1.15 / OTP 25 and latest versions with Dialyzer and lint.
• Focused retry/fault tests: 34 passed.
• Warnings-as-errors compilation, formatting, spelling, Dialyzer, and git diff --check passed.

Fixes #442.
Fixes #450.
Related: #304, #444, and Mint #486.