fix(firestore): retry transient connection errors in streaming RPCs

[bhshkh]

Retry transient connection errors (specifically syscall.ECONNRESET and syscall.ECONNREFUSED) in Firestore streaming RPCs.

Fixes: #10350

[bhshkh]

Currently, unary RPCs automatically retry on Unavailable errors (which gRPC maps connection resets/refusals to) because of default gax configs. However, streaming RPCs do not automatically retry failures that occur during stream consumption (Recv()).

This PR implements retry and resumption logic for the following streaming RPCs:

1. BatchGetDocuments (used in GetAll and DocumentRef.Get):
   • Retries the stream on transient errors.
   • Tracks remaining documents to only request what has not been successfully received yet.
   • Added strict duplicate detection in the Recv() loop to fail if the same document is received multiple times.
2. RunQuery (used in Query.Documents iteration):
   • Retries the stream on transient errors.
   • Resumes the query using StartAfter(lastDoc) if some documents were already returned to the user.
   • Adjusts the limit and clears the offset for the resumed query.
   • Resets backoff state on successful reads.
3. RunAggregationQuery (used in AggregationQuery.GetResponse):
   • Retries the stream on transient errors.
   • Since it returns a single accumulated result, it retries the entire query from the beginning.

Other changes:

• Introduced a unified isRetryable helper function in client.go that checks gRPC status codes, explicit syscall errors, and falls back to string matching ("connection reset by peer", "connection refused") for resilience.
• Refactored isPermanentWatchError in watch.go to use the new isRetryable helper.
• Updated existing tests and added new unit tests in client_test.go, query_test.go, and transaction_test.go to verify the retry behaviors.

Retries are skipped if executing within a transaction (tid != nil or tx != nil), or if ExplainOptions are set (to avoid corrupting metrics), or if LimitToLast is used on queries (due to complexity of resumption).

[gemini-code-assist]

Code Review

This pull request implements retry logic for several Firestore operations, including BatchGetDocuments, RunQuery, and RunAggregationQuery. It introduces a centralized isRetryable helper function to identify transient errors—such as specific gRPC status codes and connection resets—and utilizes exponential backoff for retries. The queryDocumentIterator was updated to resume streaming from the last received document upon retry. Corresponding unit tests were added to verify the retry behavior across these methods. I have no feedback to provide.