client-v2: cancelTransportRequest is lost when it lands between two retry attempts

[polyglotAI-bot]

Describe the bug

Client.cancelTransportRequest(queryId) (client-v2) has no effect when it lands in the window between two attempts of a retried operation: the operation issues the next attempt anyway and can complete successfully, so the caller's cancellation is silently dropped.

Cancellation state is kept per transport request (ConcurrentHashMap<String, TransportRequest> ongoingRequests, keyed by query id) instead of per operation, while a retried operation creates a new TransportRequest per attempt:

• Query path: between attempt i and attempt i+1 the registry still holds attempt i's already-completed request, so cancel() is applied to a dead request. Attempt i+1 then calls registerTransportReq(queryId, newRequest), which overwrites the entry with a fresh, un-cancelled request, so the requestIsNotCancelled(queryId) retry guard returns true and the loop keeps retrying.
• Insert paths (POJO and stream): the entry is unregistered at the end of every attempt, so in the same window ongoingRequests.get(queryId) returns null and cancelTransportRequest is a complete no-op (requestIsNotCancelled also returns true for a missing entry).

For the stream-insert path this window is reachable deterministically through public API: the client calls DataStreamWriter#onRetry() exactly between two attempts, so a cancelTransportRequest issued from a callback, another thread, or an operation-timeout watchdog that happens to fire there is lost.

Steps to reproduce

1. Point a client at a mocked server whose first response is a retryable error (HTTP 503 + X-ClickHouse-Exception-Code: 202) and whose second response is 200.
2. Insert through Client#insert(String, DataStreamWriter, ClickHouseFormat, InsertSettings) with an explicit query id and a writer whose onRetry() calls client.cancelTransportRequest(queryId).
3. The insert completes successfully and the mocked server records two requests.

Expected behaviour

Cancellation belongs to the operation, not to a single transport request: once an operation is cancelled it stays cancelled for every following attempt, so the retry loop stops instead of issuing another request and the operation fails.

Code example

String queryId = "cancel-between-attempts";
DataStreamWriter writer = new DataStreamWriter() {
    @Override public void onOutput(OutputStream out) throws IOException { out.write("1\n".getBytes()); }
    @Override public void onRetry() { client.cancelTransportRequest(queryId); }
};
// completes successfully today; a second attempt is sent after the cancel
client.insert("t", writer, ClickHouseFormat.TSV, new InsertSettings().setQueryId(queryId)).get();

Configuration

• Client version: 0.11.0-rc1-SNAPSHOT (main), pre-existing on earlier versions
• Module: client-v2 (Client.java, all three retry loops: query, POJO insert, stream insert)