Currently, in the case of a failed write op, we do not store the BulkWriteReplyItem, but store a subset of the information from the reply item in the form of a WriteError.
We then later turn it back into a BulkWriteReplyItem as we build the response items.

In cases where a write succeeds on some shards but fails on others, this causes us to lose information from the shards where the write did succeed and leads to incorrect results.
We should update our logic so that for bulkWrite, the final result (either success or failure) is always stored as a BulkWriteReplyItem, and ensure that item is a merged version of any success replies and error replies for the given write.