style: cargo fmt --workspace (whitespace/wrapping only, no semantic change)

Whole-workspace rustfmt pass picked up while iterating on Mamba GPU
backward work. Verified formatting-only via diff sampling; no logic
changed.

Co-Authored-By: Claude Sonnet 5 <[email protected]>

crates/training/rtx-transformers/src/layers/sparse_attention.rs

@@ -257,7 +257,11 @@ impl SparseAttentionLayer {
     pub fn new(config: SparseAttentionConfig, head_dim: usize) -> Self {
         let scale = 1.0 / (head_dim as f32).sqrt();
         let mask = SparseAttentionMask::new(config.clone());
-        Self { config, mask, scale }
+        Self {
+            config,
+            mask,
+            scale,
+        }
     }
 
     /// CPU reference forward pass.
@@ -581,7 +585,10 @@ mod tests {
         for i in 0..16 {
             let cnt = mask.attention_count(i);
             assert!(cnt > 0, "token {i} must attend to at least one position");
-            assert!(cnt <= 16, "token {i} cannot attend to more than seq_len positions");
+            assert!(
+                cnt <= 16,
+                "token {i} cannot attend to more than seq_len positions"
+            );
         }
     }
 
@@ -600,7 +607,10 @@ mod tests {
         };
         let mask = SparseAttentionMask::new_with_seed(config, 0);
         for i in 0..10 {
-            assert!(mask.mask[i][i], "self-attention must always be set (window_radius=0)");
+            assert!(
+                mask.mask[i][i],
+                "self-attention must always be set (window_radius=0)"
+            );
         }
     }