Keep HashCodeBuilder.append(Object) from poisoning reflection hashCode

[alhudz]

Repro: new HashCodeBuilder(17, 37).append((Object) x).toHashCode() where x.hashCode() is HashCodeBuilder.reflectionHashCode(x) returns 17 * 37 + 17 instead of 17 * 37 + x.hashCode(). The same collapse hits an ordinary acyclic graph: reflectionHashCode(holder) drops the content of any field whose own hashCode() is reflection based, so two holders differing only in such a field share a hash while reflectionEquals reports them unequal.
Cause: the cycle guard added in #1650 has append(Object) call register(object) on the same REGISTRY that reflectionAppend consults, so x's reflective hashCode() finds x already registered, mistakes it for a cycle and returns the bare initial value.
Fix: append(Object) keeps a separate append registry for its re-entrancy guard and still checks REGISTRY, so reflection cycles (including the mutually referential testReflectionObjectCycle) behave exactly as before.

Added testAppendObjectReflectionHashCodeNotPoisoned; it fails before (expected 1300, was 646) and passes after.