Fix Transcript.exons crash when GTF lacks exon_id attribute

[iskandr]

Summary

• Some GTFs (Ensembl release 54 and earlier, plus non-Ensembl GTFs) omit the exon_id attribute. pyensembl's installer already treats that column as optional (see `database.py:134`), but `Transcript.exons` still unconditionally SELECTed exon_id, so any call on such a genome crashed with:

  sqlite3.OperationalError: no such column: exon_id
  

  This was hit from pirlygenes: FN1 tests call `transcript.exons` and the local pyensembl cache (old release) has the exon table without the exon_id column.

• `Transcript.exons` now checks `db.column_exists("exon", "exon_id")` and falls back to constructing Exon objects directly from the exon row, with a synthesized per-transcript ID of the form `"<transcript_id>exon"`.

• Exon objects returned from the fallback path carry the real contig/start/end/strand/gene coordinates from the GTF; only the id is synthetic.

• Regression test builds a minimal Ensembl-style GTF with `exon_number` but no `exon_id` and verifies exon ordering and synthesized IDs.

• Bumps to 2.6.7.

Test plan

• New `test_transcript_exons_without_exon_id` passes.
• Full local suite (120 tests excluding HLA-A data-drift ones) passes.
• CI green on Python 3.9–3.12 (exercises the normal path — release 75/77/93 GTFs all have exon_id).

[coveralls]

coverage: 83.468% (+0.3%) from 83.208% — fix-exon-id-missing into main