Fix multi-node append duplicating existing chunks in index.json

[eeshsaxena]

Fixes #865.

When you run optimize(mode="append") across more than one node, the existing dataset chunks end up repeated once per node in the final index.json.

The cause is in the two-stage merge. Each node calls _merge_no_wait(node_rank, existing_index) in _done, and _merge_no_wait prepends existing_index["chunks"] to that node's {node_rank}-index.json. So every node's index file already contains the existing chunks. The last node then merges all of those per-node files together, and the existing chunks come along once for each node:

existing [A, B], node 0 adds [C], node 1 adds [D]
  node 0 -> 0-index.json = [A, B, C]
  node 1 -> 1-index.json = [A, B, D]
  final merge -> [A, B, C, A, B, D]   # wanted [A, B, C, D]

The fix adds the existing index only once. The single-node path is unchanged: it still folds the existing index in at the node-level merge. For multiple nodes, the per-node merges now carry only their own new chunks, and the existing index is folded in at the final cross-node merge instead:

  node 0 -> 0-index.json = [C]
  node 1 -> 1-index.json = [D]
  final merge (with existing) -> [A, B, C, D]

I added a test in tests/streaming/test_writer.py that walks both patterns: it shows the old per-node folding produces [A, B, C, A, B, D], and the new final-merge folding produces [A, B, C, D]. It passes locally.