[oil-native] Turn on GC in osh_eval, by lowering GC threshold!

Start with 100K objects, which seems reasonable from watching CPython's
configure run.

benchmarks/gc: Measure the cxx-opt binary, not cxx-gcverbose!

The max pause is ~141 ms, not 400+ ms.

benchmarks/gc.sh

@@ -311,8 +311,8 @@ gc-run-oil() {
 }
 
 gc-run-big() {
-  local target=_bin/cxx-gcverbose/osh_eval
-  #local target=_bin/cxx-opt/osh_eval
+  #local target=_bin/cxx-gcverbose/osh_eval
+  local target=_bin/cxx-opt/osh_eval
   ninja $target
 
   local osh=$REPO_ROOT/$target

build/ninja-rules-py.sh

@@ -43,8 +43,8 @@ int main(int argc, char **argv) {
 
   // complain_loudly_on_segfault();
 
-  // So big that we don't trigger collection
-  gHeap.Init(400 << 20);
+  // Arbitrary threshold of 100K objects based on CPython's configure
+  gHeap.Init(100000);
 
   // NOTE(Jesse): Turn off buffered IO
   setvbuf(stdout, 0, _IONBF, 0);