[CAS] Fix an off-by-one error in CAS validation

Fix a false positive in invalid CAS detection that when the last data
stored in CAS just fits inside the DataPool, validation is erroneously
think the data stored pass the end of the storage.

UnitTest for the edge case is very hard to contruct, where we might need
to insert a lot of small and different objects into CAS to be just fit
inside the CAS size. The bug was found by fuzzing the CAS, and can rely
on basic fuzzing to find regressions in the future.

rdar://150386374

Author: cachemeifyoucan

llvm/lib/CAS/OnDiskGraphDB.cpp

@@ -966,7 +966,7 @@ Error OnDiskGraphDB::validate(bool Deep, HashingFuncT Hasher) const {
       llvm_unreachable("already handled");
     case TrieRecord::StorageKind::DataPool: {
       auto DataRecord = DataRecordHandle::get(DataPool.beginData(D.Offset));
-      if (DataRecord.getTotalSize() + D.Offset.get() >= DataPool.size())
+      if (DataRecord.getTotalSize() + D.Offset.get() > DataPool.size())
         return dataError("data record span passed the end of the data pool");
       for (auto InternRef : DataRecord.getRefs()) {
         auto Index = getIndexProxyFromRef(InternRef);