feat: structured load statistics from insert_into_target_tables

[martinv13]

Summary

insert_into_target_tables previously returned a bare int (total inserted rows).
It now returns a LoadStats dataclass with inserted/existing row counts and
per-phase timing, giving callers visibility into both deduplication effectiveness
and load performance without any extra instrumentation.

Changes

Document.insert_into_target_tables → LoadStats

stats = doc.insert_into_target_tables()
print(stats.inserted)            # rows written to target tables
print(stats.existing)            # rows skipped (hash-deduplicated or parent was existing)
print(stats.duration_temp_insert)  # seconds — staging phase
print(stats.duration_merge)        # seconds — merge phase
print(stats.duration_cleanup)      # seconds — temp-table cleanup

Document.merge_into_target_tables → MergeStats(inserted, existing, duration)

Document.insert_into_temp_tables → float (phase duration)


**`Document.insert_into_target_tables`** → `LoadStats`

```python
stats = doc.insert_into_target_tables()
print(stats.inserted)            # rows written to target tables
print(stats.existing)            # rows skipped (hash-deduplicated or parent was existing)
print(stats.duration_temp_insert)  # seconds — staging phase
print(stats.duration_merge)        # seconds — merge phase
print(stats.duration_cleanup)      # seconds — temp-table cleanup

Document.merge_into_target_tables → MergeStats(inserted, existing, duration)

Document.insert_into_temp_tables → float (phase duration)