predict_pairs: add skip_unsupported option so one unsupported allele doesn't fail the whole batch

[iskandr]

BaseCommandlinePredictor.predict_pairs (added in #262) raises UnsupportedAllele for the whole batch if any allele in the input is unsupported — _resolve_supported_allele_cli_names collects the unsupported ones and raises (base_commandline_predictor.py:415).

For row-wise paired scoring where the alleles come from a fixed dataset, that's often too strict. A single allele the tool doesn't cover — e.g. NetMHCstabpan-1.0 (2016) against an allele deposited since — fails the entire call and discards every score for the alleles it does support. Callers then fall back to the exact per-allele try/except loop that predict_pairs was meant to replace:

scores = {}
for allele, rows in df.groupby("allele"):
    try:
        pred = NetMHCstabpan(alleles=[allele]).predict_peptides_dataframe(rows.peptide.unique())
    except UnsupportedAllele:
        continue   # score what the tool covers, skip the rest
    ...

Request

Add skip_unsupported=False to predict_pairs / predict_pairs_dataframe. When True, drop pairs whose allele is unsupported (omit them from the result, or return NaN for those rows) instead of raising, and surface which alleles were skipped — return them, or log at WARNING. This mirrors the graceful per-allele handling callers already do by hand, and lets predict_pairs replace the loop without changing "score what the tool can, skip the rest" into "fail the whole batch on one gap."