Store ImtIndex in ArtMethod.

This avoids recalculation and reduces pressure on the thread local cache.

This halves the time we spend hashing from 2% to 1% (maps on device).

Test: ./art/test.py -b --host --64
Change-Id: I2407bd9c222de4ddc6eea938908a1ac6d7abc35b
diff --git a/runtime/jit/jit_code_cache.cc b/runtime/jit/jit_code_cache.cc
index ff39a66..28978c5 100644
--- a/runtime/jit/jit_code_cache.cc
+++ b/runtime/jit/jit_code_cache.cc
@@ -892,7 +892,8 @@
   }
 }
 
-static void ClearMethodCounter(ArtMethod* method, bool was_warm) {
+static void ClearMethodCounter(ArtMethod* method, bool was_warm)
+    REQUIRES_SHARED(Locks::mutator_lock_) {
   if (was_warm) {
     method->SetPreviouslyWarm();
   }
@@ -1128,7 +1129,7 @@
     return false;
   }
 
-  method->ClearCounter();
+  method->SetCounter(0);
   Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
       method, GetQuickToInterpreterBridge());
   VLOG(jit)