fix(tinytorch): add missing exports and fix benchmark tests

- Module 19: Add #| export to import block so dataclass is exported
- Fix benchmark tests to use correct Benchmark API (requires models/datasets)

tinytorch/tests/19_benchmarking/test_benchmark_core.py

@@ -134,11 +134,14 @@ class TestBenchmarkMetrics:
         
         model = SimpleModel()
         x = Tensor(np.random.randn(10))
+        datasets = [[(x, None)]]
         
-        bench = Benchmark()
-        latency = bench.measure_latency(model, x)
+        bench = Benchmark([model], datasets)
+        results = bench.run_latency_benchmark(input_shape=(10,))
         
-        assert latency > 0, "Latency must be positive"
+        assert len(results) > 0, "Should produce results"
+        for name, result in results.items():
+            assert result.mean > 0, "Latency must be positive"
     
     def test_multiple_runs_are_consistent(self):
         """
@@ -161,24 +164,21 @@ class TestBenchmarkMetrics:
         
         model = SimpleModel()
         x = Tensor(np.random.randn(1, 10))
+        datasets = [[(x, None)]]
         
-        bench = Benchmark()
+        bench = Benchmark([model], datasets, measurement_runs=10)
+        results = bench.run_latency_benchmark(input_shape=(1, 10))
         
-        # Run 3 times
-        latencies = [
-            bench.measure_latency(model, x, iterations=10)
-            for _ in range(3)
-        ]
-        
-        # Check variance is reasonable (within 2x of each other)
-        max_latency = max(latencies)
-        min_latency = min(latencies)
-        
-        assert max_latency < min_latency * 3, (
-            f"Benchmark results too variable!\n"
-            f"  Latencies: {latencies}\n"
-            "Results should be within 3x of each other."
-        )
+        # Check that we get results with reasonable variance
+        for name, result in results.items():
+            # Coefficient of variation should be reasonable (std/mean < 100%)
+            if result.mean > 0:
+                cv = result.std / result.mean
+                assert cv < 1.0, (
+                    f"Benchmark results too variable!\n"
+                    f"  Mean: {result.mean}, Std: {result.std}, CV: {cv}\n"
+                    "Coefficient of variation should be < 100%."
+                )
 
 
 if __name__ == "__main__":