fix(perf): use single polars-bio overlap (no xprod) in gvl.Table

[d-laub]

Summary

• Replace per-sample polars-bio loop / pyranges cross-product in Table.count_intervals and Table._intervals_from_offsets with a single pb.overlap(queries, contig_subset_filtered_to_samples) call. Drops pyranges as a dependency entirely; removes the Python ≥ 3.12 / pyranges1 version gate.
• Bench (experiments/bench_table_overlap, 5 backends × 3 sizes × 20 trials): polars_bio no_xprod is fastest at every scale (0.173 s on 2k×200×500 vs 0.422 s for the per-sample loop, 1.543 s for pyranges1 xprod) and uses ~half the memory of either pyranges variant.
• Implementation polish: vectorized intra-cell rank (np.arange - np.repeat(boundaries, counts)), pl.Series.replace_strict for sample→index mapping, np.lexsort instead of a polars sort round-trip, hoisted numpy to module scope.

Test Plan

• `pixi run -e dev pytest tests/test_table.py -v` — 15 pass (incl. `test_table_count_intervals_matches_brute_force` and `test_table_intervals_from_offsets_roundtrip`)
• `pixi run -e dev pytest tests/dataset/test_write_tracks.py -v` — 4 pass
• `pixi run -e dev ruff check python/genvarloader/_table.py` — clean
• Optional: `pixi run -e py312 pytest tests/test_table.py -v` to confirm polars-bio resolves on py312 (was previously gated to pyranges1)

🤖 Generated with Claude Code