[python] Fix read_entries_parallel to use order-dependent merge via FileEntry.merge_entries

[plusplusjiajia]

Purpose

Fix incorrect manifest entry merge logic in read_entries_parallel() that has existed since #6451.

The old implementation collected all DELETE identifiers into a global set, then filtered out all matching ADDs. This caused two bugs:

1. Compaction file loss: When a file goes through ADD → DELETE → re-ADD (same identifier), the global DELETE set filters out both ADDs, causing the re-added file to disappear from the result.
2. Unmatched DELETE silently dropped: DELETE entries without a matching ADD were never included in the output, losing deletion information for downstream consumers (e.g., manifest compaction rewrite,
   changelog reads).