fix: mamba2_backward used grad_buf (params) not bw_d_h_s2 (trunk activation gradient)

mamba2_scan_backward kernel reads d_h_enriched [B, SH2] but was passed
self.grad_buf [TOTAL_PARAMS] — wrong buffer, wrong size. At batch_size=4096
the kernel read 1M floats from a 582K buffer → 2749 OOB reads.
Fixed: use self.bw_d_h_s2 [B, SH2] which is the actual trunk activation
gradient from the cuBLAS backward pass.

Also increased smoke test batch_size to 4096 to catch scale-dependent OOB.

compute-sanitizer: 0 errors at batch_size=4096.

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

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -2276,7 +2276,7 @@ impl GpuDqnTrainer {
             self.stream.launch_builder(&self.mamba2_backward_kernel)
                 .arg(&self.mamba2_h_history)
                 .arg(&self.save_h_s2)
-                .arg(&self.grad_buf)  // d_h_enriched = trunk gradient
+                .arg(&self.bw_d_h_s2)  // d_h_enriched = trunk activation gradient [B, SH2]
                 .arg(&w_a_ptr)
                 .arg(&w_b_ptr)
                 .arg(&w_c_ptr)

crates/ml/src/trainers/dqn/smoke_tests/generalization.rs

@@ -35,6 +35,8 @@ fn test_generalization_components_smoke() -> anyhow::Result<()> {
 
     let mut params = smoke_params();
     params.epochs = 3;
+    params.batch_size = 4096;  // Larger batch to catch OOB that only trigger at scale
+    params.buffer_size = 8192;
     params.weight_decay = 1e-4;  // G1: AdamW active
 
     let mut trainer = smoke_trainer_with(params)?;