[ohmayr]

Long ReadRows streams can drop on transient gRPC errors, which failed the whole read even though the Storage Read API can resume from any row. This reopens the stream at the next unread row.

• Retries UNAVAILABLE, and INTERNAL errors caused by RST_STREAM, up to 3 times in a row. Any received batch resets the count. DEADLINE_EXCEEDED isn't retried, since it means the caller's deadline ran out.
• An abandoned stream loses only our own listeners and keeps a no-op error listener, since google-gax emits CANCELLED after cancel().
• Errors that aren't retryable, or persist after 3 retries, destroy the stream after the batches already yielded.

[gemini-code-assist]

Code Review

This pull request introduces automatic stream resumption for BigQuery's Arrow stream reader (createQueryArrowStream) by retrying transient gRPC errors (such as UNAVAILABLE, DEADLINE_EXCEEDED, and INTERNAL with RST_STREAM) from the last successfully read offset up to a maximum of three times. Comprehensive unit tests are also added to verify this behavior. The review feedback highlights a critical issue regarding the use of removeAllListeners() on the gRPC stream during cancellation and retries, which can inadvertently strip internal library listeners and cause resource leaks. It is recommended to avoid removeAllListeners() and instead explicitly detach only the registered listeners (data, error, end) using named functions.

[ohmayr]

Rebased on the #9598–#9601 changes. Same behavior and retry rules. The retry is now a short async loop in streamReadRows around a single ReadRows attempt (readRowsOnce), so each failed attempt goes through the same listener cleanup as a normal finish, and the next attempt resumes at the next unread row. The tests check that every abandoned stream keeps none of our listeners.