fix: scope event tracking disable to CUDA Graph capture only

disable_event_tracking() was called globally in GpuDqnTrainer::new(),
polluting the CUDA context for ALL subsequent operations including
other models (TFT, PPO). This caused test_tft_adapter_deterministic
to fail when run after DQN tests.

Fix: disable/enable only inside capture_training_graph(), not at
construction. The capture region is the only place where CudaEvents
are disallowed.

Result: 855/855 ml lib tests pass (was 854/855 with 1 flaky failure).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -403,9 +403,9 @@ impl GpuDqnTrainer {
         let b = config.batch_size;
         let total_params = compute_total_params(&config);
 
-        // Disable event tracking — required for CUDA Graph capture (no CudaEvents
-        // inside captured regions). Safe: single stream, explicit ordering.
-        unsafe { stream.context().disable_event_tracking(); }
+        // Event tracking: kept ENABLED during normal ops (buffer allocation, DtoD).
+        // DISABLED only during CUDA Graph capture (in capture_training_graph).
+        // Global disable pollutes the context for other tests/models.
 
         // Stack size for the training kernel: ~46KB/thread with NUM_ATOMS=51,
         // ~12KB/thread with NUM_ATOMS=11. DIST_SIZE(SCRATCH1_DIM=256) * 4 = 11KB