feat: add CancellationToken with collect/executeStream overloads

[LantaoJin]

Which issue does this PR close?

• Closes feat: query cancellation via CancellationToken on SessionContext #68 .

Rationale for this change

A long-running DataFrame.collect(allocator) or DataFrame.executeStream(allocator) call blocks the calling Java thread for the full duration of the query. Thread.interrupt() is a no-op (the thread is parked inside runtime().block_on(...)), and there is no way to free the in-flight query's native resources early. For any embedder running request timeouts, user-cancel actions, or node-shutdown drains, this is a hard operational gap.

What changes are included in this PR?

The cancellation we ship lives on the session, not on the DataFrame. A token is allocated by the session, passed to collect / executeStream, and fired from any thread. We deliberately do not add DataFrame.cancel(): a DataFrame is a lazy plan that can be executed concurrently from multiple threads, so a per-DataFrame cancel verb has ambiguous semantics. The token is the primitive; a Spark-style ctx.addTag(name) / ctx.cancelTag(name) sugar layer can land in a follow-up.

Native shape: tokio_util::sync::CancellationToken per token, owned by a process-global registry keyed by an opaque u64 ID. JNI handlers look up by ID under a Mutex<HashMap> and clone the Arc out before any blocking work. The cloned Arc keeps the inner token alive for the borrow's lifetime, so an in-flight collect() future that already holds a clone keeps working through a concurrent close. close() removes the registry entry; the underlying token drops only when the last Arc clone goes away.

Race safety: CancellationToken.nativeHandle is final AtomicLong. cancel() / isCancelled() read via get(); close() uses getAndSet(0L) so only the winning thread issues closeToken. Combined with the registry on the native side, a concurrent close + cancel + query-start race produces at worst a clean IllegalStateException from a missing-ID lookup -- never a use-after-free of a freed Box. The registry pattern is the same scaffolding upstream issue #40 calls for across all handle types; cancellation tokens get it first because they are designed to be fired from a thread that does not own them, so the race window is the widest of any handle.

Are these changes tested?

Yes -- 19 new tests across CancellationTokenTest and DataFrameCancellationTest.

Are there any user-facing changes?

Yes -- purely additive. New CancellationToken class, one new method on SessionContext, two new overloads on DataFrame. No API removals, no deprecations, no behaviour change for existing callers. New Cargo dependency: tokio-util = { version = "0.7", features = ["rt"] }; already a transitive dep, so no new crate in the build, just a feature flag added.