Don't gc in write benchmarks

This makes it easier to evaluate the claim that "point writes are
comparable to point reads" in performance, which should be the case.

Bench.cpp

@@ -192,7 +192,6 @@ void benchConflictSet() {
     bench.run("point writes", [&]() {
       auto v = ++version;
       cs.addWrites(writes.data(), writes.size(), v);
-      cs.setOldestVersion(version - kMvccWindow);
     });
   }
 
@@ -213,7 +212,6 @@ void benchConflictSet() {
     bench.run("prefix writes", [&]() {
       auto v = ++version;
       cs.addWrites(writes.data(), writes.size(), v);
-      cs.setOldestVersion(version - kMvccWindow);
     });
   }
 
@@ -234,12 +232,13 @@ void benchConflictSet() {
     bench.run("range writes", [&]() {
       auto v = ++version;
       cs.addWrites(writes.data(), writes.size(), v);
-      cs.setOldestVersion(version - kMvccWindow);
     });
   }
 
   bench.batch(1);
 
+  cs.setOldestVersion(version - kMvccWindow);
+
   {
     bench.run("monotonic increasing point writes", [&]() {
       auto v = ++version;