Double-delivery via _process_again window between ErrorResponse and ReadyForQuery

[SeanTAllen]

After the query-in-flight sub-state machine refactoring (#67), the ReadyForQuery(non-idle) double-delivery is fixed. However, a separate double-delivery window remains due to _process_again.

_ResponseMessageParser processes one message per call, then schedules s._process_again() (a behavior call) to process the next buffered message. This creates a scheduling window between consecutive messages where other behaviors can execute:

1. ErrorResponse arrives — on_error_response delivers the error to the receiver
2. _process_again is scheduled as a behavior call
3. close() behavior executes before _process_again — on_shutdown iterates the queue and sends SessionClosed to the same receiver (double delivery)
4. _process_again would have processed ReadyForQuery and dequeued, but it's too late

This is a pre-existing issue (present before #67) and affects any pair of messages where the first delivers a result/error and the second dequeues the query.

[SeanTAllen]

Four approaches considered:

1. Track delivery in the in-flight state: Add a flag to _SimpleQueryInFlight that marks whether a result/error was delivered for the head-of-queue query. _SessionLoggedIn.on_shutdown checks it and skips the first entry if already delivered. Simple but adds a boolean to a design that just eliminated booleans.

2. Dequeue at delivery time, not at ReadyForQuery: Won't work cleanly — multi-statement queries get multiple CommandComplete messages before one ReadyForQuery, so you can't dequeue on the first CommandComplete.

3. Process all remaining messages synchronously: Remove _process_again and loop until the buffer is empty or incomplete. Eliminates the window entirely but risks starving other actors on large result sets.

4. Hybrid: process synchronously within a query cycle (favored): Loop synchronously until ReadyForQuery is seen or the buffer is exhausted, then yield. This scopes the synchronous processing to a single query's response, eliminating the interleaving window while still yielding between query cycles. If a callback triggers shutdown during the loop, on_shutdown clears the read buffer, causing the next parse to return None and exit the loop naturally.

PR #69 implements approach 4.