fix: predictCoding on empty ranges returns AAStringSet for REFAA/VARAA (#86)

[jmg421]

Problem

When predictCoding() is called with a query that has no overlap with any CDS (e.g. a non-coding variant), .localCoordinates() returns a zero-length GRanges. The early-exit guard at that point returned txlocal directly — before REFAA/VARAA columns were ever added to mcols(). This caused downstream operations like reverse() or subseq() on those columns to throw errors.

Reproducer from #86:

library(VariantAnnotation)
library(BSgenome.Hsapiens.UCSC.hg19)
library(TxDb.Hsapiens.UCSC.hg19.knownGene)

fl <- system.file("extdata", "chr22.vcf.gz", package="VariantAnnotation")
vcf <- readVcf(fl, "hg19")[123]  # non-coding variant
seqlevels(vcf) <- "chr22"
coding <- predictCoding(vcf, TxDb.Hsapiens.UCSC.hg19.knownGene, seqSource=Hsapiens)

coding$REFAA  # NULL — should be AAStringSet()
coding$VARAA  # NULL — should be AAStringSet()

Fix

Two changes in R/methods-predictCoding.R:

1. Remove the early return on length(txlocal) == 0 — let execution fall through to the full mcols()-building block, which naturally produces zero-length AAStringSet columns via AAStringSet(rep("", length(txlocal))).

2. Fix scalar GENEID — change GENEID=NA_character_ to GENEID=rep(NA_character_, length(txlocal)) so DataFrame() construction is valid at zero length.

Test

Extended test_predictCoding_empty in inst/unitTests/test_predictCoding-methods.R to assert:

• mcols(result)$REFAA is an AAStringSet
• mcols(result)$VARAA is an AAStringSet
• Both have length == 0L

Fixes #86.