fix(aws_s3 source): complete in-flight SQS message processing before shutdown

[sanjams2]

Problem

During shutdown, tokio::select! in IngestorProcess::run() cancels run_once() mid-flight, which can leave SQS messages undeleted after events have already been sent to sinks, causing duplicate delivery after visibility_timeout expires.

Fix

Split run_once() into two distinct phases:

1. Cancellable receive_with_backoff() — polls SQS for messages with exponential backoff on errors. This is raced against shutdown in select! and is safe to cancel because no messages have been processed yet.

2. Non-cancellable process_messages() — handles each received message, forwards deferred entries, and batch-deletes successfully processed messages from SQS. This runs to completion without being raced against shutdown, ensuring SQS deletes are never skipped.

This preserves all existing behavior (error metrics, deferred queue forwarding, batch delete handling, backoff logic) while eliminating the race condition.

Tests

Added 3 new unit tests:

• shutdown_interrupts_long_poll — verifies that shutdown interrupts the SQS receive phase cleanly
• shutdown_interrupts_backoff_sleep — verifies that shutdown interrupts the backoff sleep between retries
• inflight_processing_completes_despite_shutdown — verifies that in-flight message processing (including SQS delete) runs to completion even when shutdown fires mid-processing

Closes #24010