Pre-train and Search: Efficient Embedding Table Sharding with Pre-trained Neural Cost Models