[alejandrogzi]

Adds SpliceAI2 to core: four modules, a subworkflow and a workflow, plus the catalog entry.

What is added

• modules/nextflow/spliceai2/chunk: splits the genome into strand-aware chunks (spliceai2 chunk).
• modules/nextflow/spliceai2/predict: per-chunk donor and acceptor WIGs on both strands (spliceai2 predict, GPU).
• modules/nextflow/spliceai2/transcripts: optional transcripts BED12 for a loci BED6 (spliceai2 transcripts, GPU).
• modules/nextflow/spliceai2/publish: copies the final outputs.
• subworkflows/spliceai2: chunk → predict → WIG to bigWig → per-track merge, and the optional transcripts branch.
• workflows/spliceai2: the entry point (--genome, --spliceai2_models, optional --spliceai2_loci).

Usage notes

• The container is ghcr.io/hillerlab/spliceai2:latest (0.0.6, from Add SpliceAI2 container: genome-wide splice sites and per-locus transcripts containers#69).
• The model weights are not in the image. Pass --spliceai2_models pointing to a directory with
  spliceai2_multispecies_13M_01/ and spliceai2_multispecies_13M_02/, obtained under Illumina's terms.
• Inference runs with --no-compile: torch.compile needs a writable HOME and a C compiler, and the Apptainer task has neither.
  Other flags can still be passed through ext.args.
• --spliceai2_loci needs a FASTA index next to the genome: genome.fa.fai, or for genome.fa.gz an index of the uncompressed
  sequence named genome.fa.fai. Core has no samtools faidx module, so a missing index fails early with a message saying so.

Tests

Real runs with -profile apptainer on an L4, using only hillerlab.config and the published image:

run	tasks
uncompressed FASTA + loci	13/13 exit 0
compressed chunks + loci	13/13 exit 0
no loci	12/12 exit 0
gzipped FASTA + loci	14/14 exit 0

• The workflow's four WIGs are byte-identical to a direct spliceai2 predict run.
• The merged bigWigs match the direct output at every base on both strands.
• The transcripts BED12 is byte-identical to a direct spliceai2 transcripts run, including the gzipped-genome run.
• The SpliceAI2 subworkflow stub runs pass.

Known issue (pre-existing, unchanged here)

modules/nextflow/chromsize's stub writes ${prefix}/chrom.sizes with prefix = meta.id, while the real script and the
declared output use genome.baseName. A full-workflow -stub-run therefore stops at CHROMSIZE. Real runs are unaffected.