Add a query timeout #19

Problem

A query can't be given a time limit. execute() blocks until the query finishes. The only way to stop it is Ctrl-C: block_on() releases the GIL and checks for signals every 100 ms, cancelling the query on KeyboardInterrupt. A service or a benchmark harness can't rely on that, so a query with an unexpectedly large intermediate result (see #11) runs until it finishes or the process is killed.

Evidence

Every other engine in both benchmark repositories enforces a 10-second limit per query (Neo4j's transaction timeout, Kuzu's and Ladybug's set_query_timeout, DuckDB's interrupt(), FalkorDB's per-query timeout). lance-csr is the only one that runs without a limit, which both repositories' READMEs have to note.

Suggested fix

Add a timeout argument (seconds, default none) to CypherEngine.execute(), execute_batches() and CypherQuery.execute(), enforced as a third branch of the existing tokio::select! in block_on()'s result_or_interrupt(): when the deadline passes, the query future is dropped, which cancels its tasks, and TimeoutError is raised. execute_batches() needs the deadline applied to each batch read. Check that long-running custom operators (CSR expansion, shortest path) yield to the runtime often enough to be cancelled promptly.

Acceptance criteria

• A query that exceeds timeout raises TimeoutError within a small margin, and its tasks stop (no background CPU afterwards).
• The engine stays usable after a timeout; a test runs a query after one.
• docs/api-reference.md and the relevant guide document the argument.
• The benchmark runners set it, so lance-csr follows the same rules as the other engines.

Identified from the benchmark setup; not prototyped.