perf: extend the benchmark for better resolving checks

The current schema generated for benchmarks on large schemas didn't
generate any definitions that use type references. Therefore, the
benchmark didn't give good insight on type resolution timing.

crates/stef-benches/src/lib.rs

@@ -99,5 +99,20 @@ pub fn generate_schema(count: usize) -> String {
         writeln!(&mut input, "type Alias{i} = SampleNamed;").unwrap();
     }
 
+    input.push_str("\nmod resolve {\n");
+    writeln!(&mut input, "    struct Sample {{").unwrap();
+    for i in 1..=count {
+        writeln!(&mut input, "        field{i:05}: Type{i:05} @{i},").unwrap();
+    }
+    writeln!(&mut input, "    }}").unwrap();
+    for i in 1..=count {
+        writeln!(&mut input, "    struct Type{i:05} {{").unwrap();
+        writeln!(&mut input, "        value1: string @1,").unwrap();
+        writeln!(&mut input, "        value2: i64 @2,").unwrap();
+        writeln!(&mut input, "        value3: u64 @3,").unwrap();
+        writeln!(&mut input, "    }}").unwrap();
+    }
+    input.push_str("}\n");
+
     input
 }