Fix/iceberg drop tables

[hentzthename]

Hi Nico, another small one that fell out of running pipeline.extract(source, refresh="drop_resources") after pipeline.sync_destination() against a Nessie deployment.

The load layer emitted:

Client for iceberg_rest does not implement drop table.
Following tables {'x', 'y'} will not be dropped

…and silently skipped the drops, so stale tables stuck around across refreshes.

Problem

dlt core gates the per-table drop path on hasattr(job_client, "drop_tables") (dlt/load/utils.py:170). IcebergRestClient only exposed drop_storage() (a full namespace wipe) -- no drop_tables(*names, delete_schema=True) -- so the load layer fell back to the warn-and-skip branch. Net effect:

• refresh="drop_resources" / refresh="drop_sources" were effectively no-ops on this destination.
• pipeline.sql_client().drop_dataset() had no coherent per-table partner (dataset-level works via the base-class DROP SCHEMA CASCADE).
• Consumers had to reach around the destination with pyiceberg directly for destructive ops.

---

Solution

Implement the JobClient.drop_tables contract on IcebergRestClient:

• Drop each named table via the PyIceberg catalog, swallowing NoSuchTableError so the call is idempotent (dlt may pass tables that were never physically created).
• When delete_schema=True, remove all _dlt_version rows where schema_name = self.schema.name via table.delete(EqualTo(...)), matching the SqlJobClientBase.drop_tables contract.

One deviation worth calling out: the obvious move would be self._delete_schema_in_storage(self.schema), but that method lives on SqlJobClientBase (not JobClientBase) and uses self.sql_client.execute_sql(...). IcebergRestClient extends JobClientBase directly, and its sql_client is a DuckDB view provider rather than a real DDL-capable client -- so the DELETE is issued via PyIceberg's row-delete API instead, reusing the pattern already at destination_client.py:1151-1153.

---

Changes

destination_client.py

Symbol	Change
IcebergRestClient.drop_tables (new)	Drops each named table via catalog.drop_table(...); NoSuchTableError is swallowed. When delete_schema=True, deletes _dlt_version rows for self.schema.name via version_table.delete(EqualTo("schema_name", ...)).

No changes to sql_client.py or schema_evolution.py.

---

Tests

• New test_drop_tables.py covering:
  • the hasattr gate (method is actually exposed on the class)
  • selective drop of named tables only
  • idempotent behavior on missing tables
  • delete_schema=True clears _dlt_version rows for the current schema
  • pipeline.run(..., refresh="drop_resources") end-to-end (the originally reported symptom)
• Full suite: 175/175 passing locally (SQLite + Nessie + Polaris + Lakekeeper).