feat: add scope parameter to insert-only merge strategy (#3827)

[mattiasthalen]

Description

Add an optional scope parameter to the insert-only merge strategy that restricts key matching to the most recent load instead of all history.

@dlt.resource(
    write_disposition={
        "disposition": "merge",
        "strategy": "insert-only",
        "scope": "previous_load",
    }
)

What it does:

• scope=None (default): current behavior — dedup keys against entire destination table
• scope="previous_load": dedup keys against only the most recent _dlt_load_id, allowing re-insertion of records that reappear after being absent (A→B→A pattern)

Implementation across backends:

• SQL destinations: pre-filter staging to remove keys present in previous load, then MERGE ON FALSE to insert remaining rows. Nested tables join through root table since they lack _dlt_load_id.
• Delta: _filter_by_previous_load compares source Arrow data against target rows from the previous load. Child tables resolved via root table's _dlt_id/_dlt_load_id.
• Iceberg: same approach using pyiceberg scan with EqualTo filter. Child tables resolved via root table.
• LanceDB: raises DestinationTerminalException (not supported).

Validation: invalid scope values (e.g. scope='prev_load') are rejected at resource definition time via ValueErrorWithKnownValues.

Related Issues

• Closes Snapshot-diff ingestion for append-only data warehousing #3827

Additional Context

Architecture docs (in dlt-sandbox):

• Design spec
• Implementation plan
• Merge strategy decision
• Review fixes decision

Test coverage (16 tests, all passing):

• A→B→A pattern with scope=previous_load (duckdb, delta, iceberg)
• Default scope unchanged (backward compat)
• Empty target table
• Nested tables with scope (duckdb, delta, iceberg)
• Hard delete + scope
• Invalid scope validation
• Valid scope acceptance

Tests run locally and pass before submitting.