Extended-protocol fixes: Flush handling + prepared-statement client-cache consistency

[ikeboy003]

Four commits. Two distinct fixes that share a theme: pgcat losing extended-protocol state in ways the client cannot detect or recover from.

1. Flush ('H') handling — 3 commits

Drivers using describeFirst (e.g. postgres.js with prepare: false + params) send Flush between Parse/Describe and Bind/Execute. pgcat logged Unexpected code: H and dropped it, hanging the client.

• Handle Flush ('H') extended-protocol message — drains buffered extended-protocol messages into the send buffer, forwards the Flush byte, keeps the server checked out for the follow-up Bind/Execute/Sync.
• Read Flush responses without waiting for ReadyForQuery — Server::recv() only breaks on 'Z', which the server never sends after Flush. Adds recv_flush_response() that breaks on the actual terminal markers (RowDescription / NoData / ErrorResponse / BindComplete).
• Flush response: drop double-copy parse and clone — perf: inspect framing byte without re-parsing length or copying the body for messages we only need to identify.

2. Prepared-statement client-cache desync — 1 commit

ensure_prepared_statement_is_on_server on PreparedStatementError:

1. Removed the entry from the client's prepared_statements map
2. Returned Ok(())

The client never received Close('S') so it still believed the statement was valid. The next Bind for that name missed in buffer_bind and pgcat dropped the entire TCP conn with ClientError(\"Prepared statement sN doesn't exist\").

For drivers that cache prepared statements per logical connection (tokio-postgres, asyncpg, pgx) and reuse them across many transactions, this is unrecoverable — the client has no way to know its view diverged from pgcat's.

Fix:

• Keep the client cache intact — it mirrors what the client believes; dropping it without Close is a protocol-level lie.
• Mark the offending server bad so the pool replaces it on next checkout.
• Propagate PreparedStatementError so the caller fails this one op cleanly and retries on a fresh backend.

Test plan

• cargo build clean
• cargo fmt --check
• postgres.js prepare: false + params: completes in ~500ms (was hanging)
• Happy to add an integration test for Flush + describeFirst if you point me at the right harness