feature: add a backwards-compatibility fixture written with stable row ids

Stable row ids are being taken from experimental to generally available, which makes their on-disk representation a durable compatibility contract. There is currently no test proving that contract holds.

test_data/ has 22 fixture directories covering format and index evolution — v0.30.0_pre_created_at for the index created_at field, fri_straddle_pre_6610 for fragment reuse, and version fixtures back to v0.5.9. None of them was written with enable_stable_row_ids=True. So there is no "an older writer produced this, a current reader must still read it correctly" test for:

• RowIdSequence and its five segment encodings (Range, RangeWithHoles, RangeWithBitmap, SortedArray, unsorted Array), including the EncodedU64Array bitpacking variants
• RowIdMeta::Inline bytes in DataFragment
• manifest.next_row_id and FLAG_STABLE_ROW_IDS
• RowDatasetVersionSequence for _row_created_at_version and _row_last_updated_at_version

The gap matters more than usual here because the encodings are chosen adaptively by a cost model, so a change to that model silently changes which encoding gets written. Nothing today would catch a reader-side regression on an encoding the current writer stopped choosing.

Suggested shape, following the existing convention:

• A test_data/<version>/datagen.py that asserts the Lance version it was generated with, creates a dataset with enable_stable_row_ids=True, and deliberately produces a spread of segment encodings — an append-only fragment (Range), one with scattered deletions (RangeWithHoles), one with dense deletions (RangeWithBitmap), and a compacted/merged fragment (SortedArray).
• Rows that have been updated, so _row_created_at_version and _row_last_updated_at_version diverge and the RLE runs are non-trivial.
• A Rust test using copy_test_data_to_tmp that opens the fixture, asserts _rowid values are exactly what the generator recorded, asserts a take by row id returns the right rows, and runs Dataset::validate().

Related: the storage layout for these sequences is expected to change (they are moving out of fragment metadata), which is precisely why a fixture written before that change should exist first.

Start by comparing the existing test_data fixture directories and their datagen.py conventions, especially v0.30.0_pre_created_at and fri_straddle_pre_6610. Then trace the Rust test pattern using copy_test_data_to_tmp. Done means a stable-row-id fixture covers the requested segment encodings and updated rows, while the reader test verifies recorded _rowid values, row-id take results, and Dataset::validate().