feat: query cancellation via `CancellationToken` on `SessionContext`

從 native/src/lib.rs 中的 JNI 阻塞點開始，檢查 SessionContext、DataFrame 和 resource-handle 的 Java 入口點。將提議的 token 生命週期和 collect/executeStream 多載與 cancellation.rs 和 query_tracker.rs 中的參照進行比較。完成的標準是：列出的 API 支援取消和清理，同時不更改現有的 zero-token 方法，並且在收集和串流期間可以觀察到取消。

Is your feature request related to a problem or challenge?

A long-running DataFrame.collect(allocator) or DataFrame.executeStream(allocator) call blocks the calling Java thread for the entire duration of the query. Thread.interrupt() does nothing — the JNI thread is parked inside runtime().block_on(...) (native/src/lib.rs), and the interrupt flag is ignored by the Tokio runtime. There is no way to abort an in-flight query, free its native resources early, or unblock the calling thread short of waiting for the query to finish.

For any embedder running multi-tenant workloads — request timeouts, user-cancel actions, node shutdown, leader-election handover — this is a hard operational gap. The OpenSearch analytics backend (OpenSearch/sandbox/plugins/analytics-backend-datafusion/rust/src/cancellation.rs and query_tracker.rs) carries a CancellationToken-based wrapper precisely because upstream offers nothing.

This is complementary to issue #40 (close()/JNI use-after-free race) but distinct: #40 is about safely tearing down a finished handle; this is about signalling an in-flight future to stop. Both eventually share the atomic-handle scaffolding from #40's option 2, so coordination is worthwhile, but the surface lands cleanly without #40 having to merge first.

Describe the solution you'd like

A token-based cancellation API on SessionContext, modeled on Spark 4.0's interruptTag shape (cancel lives on the session, not on the DataFrame). The token is a separate handle from the DataFrame so cancel can fire from a thread that does not hold the DataFrame.

v1 surface

try (SessionContext ctx = new SessionContext();
     CancellationToken token = ctx.newCancellationToken();
     DataFrame df = ctx.sql("SELECT ... FROM big_table")) {

    Future<ArrowReader> fut = pool.submit(() -> df.collect(allocator, token));

    // from another thread (timeout watcher, user-cancel handler, ...):
    token.cancel();

    // fut completes with CancellationException
}

New methods:

• SessionContext.newCancellationToken() -- returns a fresh CancellationToken bound to this session.
• CancellationToken.cancel() -- fires the token; idempotent.
• CancellationToken.isCancelled() -- non-blocking check.
• CancellationToken.close() -- releases the native handle; the token is AutoCloseable so try-with-resources handles cleanup.
• DataFrame.collect(BufferAllocator, CancellationToken) -- overload that takes a token. The existing zero-token collect(BufferAllocator) is unchanged.
• DataFrame.executeStream(BufferAllocator, CancellationToken) -- same overload pattern. Token is held by the returned ArrowReader for its full lifetime; cancel mid-stream aborts the next loadNextBatch().

Describe alternatives you've considered

No response

Additional context

Out of scope

• Tag form. Ship the token primitive first; tag is sugar that can land in a follow-up if a user actually asks for it.
• Sync-API breakage. df.collect(allocator) keeps working unchanged; the new method is df.collect(allocator, token) (overload).
• Per-operator cancel granularity. Today the cancel point is each block_on site; sub-operator cancellation is upstream-DataFusion territory.