chore: remove unnecessary black_box in divan and bencher compat layers

Apply the same optimization from PR 98 to both divan and bencher compat layers:
- Remove black_box around routine calls during measurement
- Only apply black_box to outputs after measurement ends
- Reduces stack usage and eliminates unnecessary memory operations
- Improves benchmark performance by avoiding artificial overhead

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

Author: art049

crates/bencher_compat/src/compat.rs

@@ -38,8 +38,9 @@ impl Bencher {
             black_box(inner());
         }
         self.codspeed.start_benchmark(uri.as_str());
-        black_box(inner());
+        let output = inner();
         self.codspeed.end_benchmark();
+        black_box(output);
     }
 }
 

crates/divan_compat/src/compat/bench/mod.rs

@@ -97,8 +97,9 @@ where
         let mut gen_input = self.config.gen_input.borrow_mut();
         let input = gen_input();
         codspeed.start_benchmark(self.uri.as_str());
-        divan::black_box(benched(input));
+        let output = benched(input);
         codspeed.end_benchmark();
+        divan::black_box(output);
     }
 
     pub fn bench_local_refs<O, B>(self, mut benched: B)
@@ -110,7 +111,8 @@ where
         let mut input = gen_input();
 
         codspeed.start_benchmark(self.uri.as_str());
-        divan::black_box(benched(&mut input));
+        let output = benched(&mut input);
         codspeed.end_benchmark();
+        divan::black_box(output);
     }
 }