create_missing_features: opt-in / configurable strictness for synthesized parent rows

Observation

pyensembl calls gtfparse.create_missing_features to backfill transcript and gene rows that the GTF omits (legitimate when the upstream only emits exon / CDS). The synthesizer derives the parent feature from the child rows it sees, which is usually correct — but when the input GTF deviates in ways the synthesizer doesn't anticipate, downstream consumers have to clean up.

Two recent pyensembl issues are downstream of this:

• openvax/pyensembl#331 / v2.6.4 — older Ensembl releases (e.g. release 54) omit exon_id. create_missing_features synthesizes transcript rows that don't carry exon_id either, and pyensembl's Transcript.exons then crashed with sqlite3.OperationalError: no such column: exon_id. Workaround lives in pyensembl: a column_exists("exon", "exon_id") check at transcript.py:149.
• openvax/pyensembl#252 / v2.6.4 — TAIR chr_patch_hapl_scaff fragments contain CDS rows without matching start_codon/stop_codon. After feature synthesis the transcript row exists but coding_sequence couldn't compute its endpoints. Workaround in pyensembl: Transcript.coding_sequence returns None rather than raising.

In both cases the inconsistency is structural to the input GTF rather than to gtfparse's logic, but a caller-side knob would help.

Proposal

Add a strict / policy kwarg to create_missing_features:

• strict="raise" — when a child row lacks attributes the synthesizer would need to faithfully build the parent, raise.
• strict="warn" — log + skip the synthesized row.
• strict="best_effort" — current behavior; synthesize whatever we can, fill missing attrs with None.

Or more granularly: a required_attributes_per_feature: dict[str, set[str]] knob so callers can pin "if you're going to synthesize a transcript row, only do so when the child exon carries exon_id".

Why it'd help

The two issues above both look like "pyensembl handles partial GTFs" but the upstream signal that something's off (exon_id missing, start_codon missing) gets eaten by the synthesizer and lives later as a crash at query time.

Scope guard

• Existing default behavior should stay best-effort. This is purely additive.
• Bacteria/archaea GFFs with their own synthesis quirks are out of scope for this issue (separate parser ecosystem).

Start at gtfparse.create_missing_features and review the downstream examples in pyensembl's transcript.py, including the column_exists("exon", "exon_id") check and Transcript.coding_sequence behavior. Compare the proposed strict/policy options and required_attributes_per_feature design, preserving best-effort behavior by default. Done means callers can opt into raising or warning/skipping when synthesized parent attributes are insufficient, with the existing default unchanged.