Refactor merge logic to use one abstruct

[yanghua]

To track this issue: #5117 (comment)

cpoied here:

• The partition sources of the single-machine and the distributed mode are different:
  • for the single machine: the data comes from the memory;
  • for distributed: the data is persisted into the storage;
• About the file writer initatization they are also different:
  • for the single machine, it is initialized in the merge_partitions method;
  • for the distributed: we implement via init_writer_for_flat/pq/sq for different vector index type;
• About the logic of the merger, they are also different due to different partition data sources.

For different partition sources, it would be better to abstract a PartitionSource trait. After that, we can introduce a UnifiedPartitionMerger to do a general merger. Introducing a StorageWriterFactory to create different writers.

For a merger, the generic logic can be split into four common steps:

• create merger;
• instantiate merger;
• merger#merge();
• write final metadata;