Cache Registrars in memory

This replaces the memcache caching, which we think is overall a bad idea.
We load all registrars at once instead of caching each as needed, so that
the loadAllCached() methods can be cached as well, and therefore will
always produce results consistent with loadByClientIdCached()'s view of the
registrar's values. All of our prod registrars together total 300k of data
right now, so this is hardly worth optimizing further, and in any case this
will likely reduce latency even further since most requests will be
served out of memory.

While I was in the Registrar file I standardized the error messages for incorrect
password and clientId length to be the same format, and cleaned up a few
random things I noticed in the code.

-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=156151828

java/google/registry/export/SyncGroupMembersAction.java

@@ -113,13 +113,13 @@ public final class SyncGroupMembersAction implements Runnable {
    */
   @Override
   public void run() {
-    List<Registrar> dirtyRegistrars = Registrar
-        .loadAllActive()
+    List<Registrar> dirtyRegistrars = FluentIterable.from(Registrar.loadAllCached())
         .filter(new Predicate<Registrar>() {
           @Override
           public boolean apply(Registrar registrar) {
-            // Only grab registrars that require syncing and are of the correct type.
-            return registrar.getContactsRequireSyncing()
+            // Only grab active registrars that require syncing and are of the correct type.
+            return registrar.isActive()
+                && registrar.getContactsRequireSyncing()
                 && registrar.getType() == Registrar.Type.REAL;
           }})
         .toList();