Replace the two weak tests that passed on the unfixed code with proper
regression coverage:

1. test_task_to_record_batches_does_not_use_table_filter_without_positional_deletes
   Replaces pyarrow.Table with a sentinel class whose from_batches raises
   AssertionError.  A custom metaclass preserves isinstance checks so the
   rest of the code-path is unaffected.  With the fix the sentinel is never
   reached; without the fix it fires immediately, detecting the old
   pa.Table.from_batches / filter / combine_chunks / to_batches[0] path
   that caused the SIGSEGV on Apple Silicon.

2. test_task_to_record_batches_filter_applied_after_positional_deletes
   Uses data [1,2,3,4,5] with positional deletes on positions 1 and 3
   (removing values 2 and 4), then applies filter id > 2.  Expected
   result [3, 5] would be wrong if either deletes or the subsequent
   filter were skipped.

- Combine the two positional-delete handling blocks into one: apply
  take(indices) and the row filter together inside the single
  'if positional_deletes' block, removing the redundant mid-loop
  empty-batch check (filtering an empty batch is free).
- Replace the sentinel-based regression test (which passed even when
  the fix was reverted) with a behavioral test that picks a scenario
  where the old bug produces a distinct wrong answer:
    data=[1,2,3,4], pos_delete=2 (value 3), filter=id>2
    correct: [4]   old bug (scanner pre-filters): [3,4]
  The assertion on [4] will fail against any regression.

Addresses review feedback from @ebyhr and @Fokko.