fix(critical): IQN d_h_s2 is f32 not bf16 — remove bogus bf16_to_f32 cast

The IQN backward kernel writes f32 via atomicAdd into d_h_s2_buf
(allocated as CudaSlice<f32>). But apply_iqn_trunk_gradient called
bf16_to_f32_kernel on this buffer, reinterpreting f32 bits as bf16.
This produced garbage/NaN that poisoned the trunk gradient on H100.

Replace with direct f32 DtoD copy — no cast needed.

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

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -1111,26 +1111,12 @@ impl GpuDqnTrainer {
             }
         }
 
-        // ── 2. Cast IQN d_h_s2 (bf16) → bw_d_h_s2 (f32) ─────────────────
-        // IQN head produces bf16 gradient. Cast to f32 for the backward scratch.
+        // ── 2. DtoD copy IQN d_h_s2 (f32) → bw_d_h_s2 (f32) ────────────
+        // IQN d_h_s2_buf is f32 (atomicAdd in backward kernel). No cast needed.
         {
-            let src = iqn_d_h_s2_ptr;
-            let dst = self.ptrs.bw_d_h_s2;
-            let n_elems = (b * sh2) as i32;
-            let blocks = ((b * sh2 + 255) / 256) as u32;
-            unsafe {
-                self.stream
-                    .launch_builder(&self.bf16_to_f32_kernel)
-                    .arg(&src)
-                    .arg(&dst)
-                    .arg(&n_elems)
-                    .launch(LaunchConfig {
-                        grid_dim: (blocks, 1, 1),
-                        block_dim: (256, 1, 1),
-                        shared_mem_bytes: 0,
-                    })
-                    .map_err(|e| MLError::ModelError(format!("IQN d_h_s2 bf16→f32: {e}")))?;
-            }
+            let n_bytes = b * sh2 * std::mem::size_of::<f32>();
+            dtod_copy(self.ptrs.bw_d_h_s2, iqn_d_h_s2_ptr, n_bytes, &self.stream, 0, "iqn_d_h_s2_copy")
+                .map_err(|e| MLError::ModelError(format!("IQN d_h_s2 f32 DtoD: {e}")))?;
         }
 
         // ── 3. ReLU mask: bw_d_h_s2 *= (save_h_s2 > 0) (f32 dx, bf16 act) ──