Structural variant effects() is very slow at some loci and for large typed SVs

Annotating every SV in the four esvee somatic VCFs from the public osteosarc.com dataset (1,797 SVs after loading, GRCh38 / Ensembl 115) with Variant.effects() took about 21 minutes. A handful of breakends dominate:

A native stack sample during the run was dominated by SQLite (sqlite3VdbeExec, B-tree reads), i.e. pyensembl queries, rather than Python-side work.

Separately, breakend pairs that callers label SVTYPE=DEL/DUP/INV now load as typed SVs spanning the event, so effects() visits every transcript in the span. Some esvee events span tens of megabases (e.g. a 37 Mb INV on chr2), which makes full-span annotation expensive.

Ideas:

• profile StructuralVariantAnnotator.annotate_on_transcript and the cryptic-exon / splice-outcome enumeration it runs per effect;
• cache per-transcript exon and cDNA lookups across effects;
• skip splice / cryptic enumeration for transcripts wholly inside a span, where no breakpoint touches them;
• offer a breakpoint-only mode for very large SVs.

Start at StructuralVariantAnnotator.annotate_on_transcript and profile the cryptic-exon and splice-outcome enumeration, along with the pyensembl SQLite queries. Reproduce the four osteosarc.com VCF benchmark and inspect the listed slow breakends and large typed SVs. Done means establishing a faster effects() path while preserving annotation results for these cases.