fix(bf16): EventTrackingGuard for post-graph kernel launches

Root cause: cudarc event tracking left stale state after CUDA graph
replay, causing INVALID_VALUE on subsequent kernel launches.
Fix: add EventTrackingGuard to clip_grad_buf_inplace (same pattern
as read_grad_norm_sync, apply_spectral_norm, etc.).

Smoke test now runs 3 epochs / 600 steps successfully.
Q-values diverge (bf16 precision) — needs hyperparameter tuning
for bf16 (lower lr, smaller Huber delta, tighter gradient clip).

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

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -1521,7 +1521,9 @@ impl GpuDqnTrainer {
     /// from overwhelming the subsequent auxiliary gradient additions.
     /// All operations are async on the same stream — zero CPU sync.
     pub fn clip_grad_buf_inplace(&mut self, max_norm: f32) -> Result<(), MLError> {
-        // clip_grad_buf_inplace: standalone kernel launch (not captured in CUDA graph)
+        // Disable cudarc event tracking — graph replay leaves stale events
+        // that cause INVALID_VALUE on post-graph kernel launches.
+        let _evt_guard = EventTrackingGuard::new(self.stream.context());
         // Zero the grad_norm accumulator
         self.stream
             .memset_zeros(&mut self.grad_norm_buf)
@@ -3571,7 +3573,6 @@ impl GpuDqnTrainer {
     fn launch_grad_norm(&self) -> Result<(), MLError> {
         let tp = self.total_params as i32;
         let blocks = ((self.total_params + 255) / 256) as u32;
-            tp, blocks, self.ptrs.grad_buf, self.ptrs.grad_norm_buf);
 
         let launch_cfg = LaunchConfig {
             grid_dim: (blocks, 1, 1),