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-compress/src/quantization/w4a16_matmul.rs

@@ -252,8 +252,8 @@ pub fn w4a16_matmul_cpu(
                 //   next_group_in_col = next_group_flat - weight_row_offset
                 //   (clamped to in_features)
                 let next_group_flat = (group_idx + 1) * group_size;
-                let next_in_col = (next_group_flat.saturating_sub(weight_row_offset))
-                    .min(in_features);
+                let next_in_col =
+                    (next_group_flat.saturating_sub(weight_row_offset)).min(in_features);
 
                 let scale = weights.scales[group_idx];
                 let zero = weights.zeros[group_idx] as f32;
@@ -264,11 +264,7 @@ pub fn w4a16_matmul_cpu(
 
                     // Per-column activation scale (inverse of the activation-derived
                     // scale applied during AWQ quantisation).
-                    let act_scale = weights
-                        .activation_scales
-                        .get(col)
-                        .copied()
-                        .unwrap_or(1.0);
+                    let act_scale = weights.activation_scales.get(col).copied().unwrap_or(1.0);
 
                     let w_fp32 = (q - zero) * scale * act_scale;
                     let a_fp32 = activations[act_row_offset + col];
@@ -624,9 +620,7 @@ mod tests {
             ref_result.len(),
             "output length mismatch"
         );
-        for (idx, (&fused, &reference)) in
-            fused_result.iter().zip(ref_result.iter()).enumerate()
-        {
+        for (idx, (&fused, &reference)) in fused_result.iter().zip(ref_result.iter()).enumerate() {
             let diff = (fused - reference).abs();
             assert!(
                 diff <= tolerance,
@@ -652,9 +646,7 @@ mod tests {
         // out[b, 0] = 2*b + 0*b = 2b
         // out[b, 1] = 0*b + 2*b = 2b
         let w = make_weight(2, 2, vec![2, 0, 0, 2], 1.0, 0, vec![1.0; 2]);
-        let activations: Vec<f32> = (0..4)
-            .flat_map(|b| vec![b as f32, b as f32])
-            .collect();
+        let activations: Vec<f32> = (0..4).flat_map(|b| vec![b as f32, b as f32]).collect();
 
         let result = w4a16_matmul_cpu(&w, &activations, 4).unwrap();
         assert_eq!(result.len(), 8);