Add MixTCRpred model catalog and wrapper

[iskandr]

Summary

• extend TCR with backward-compatible paired-chain V/J fields and common upstream column aliases
• add a pinned, license-gated MixTCRpred source fetch plus selective checksum-verified Zenodo model downloads
• expose the full 147-target model catalog in Python and through mhctools ls mixtcrpred --models, including shipped and optional weights, paths, managers, provenance, training counts, AUCs, and confidence tier
• add fixed-pMHC prediction APIs and a CSV CLI that preserve raw score, percentile rank, corrected V/J genes, V-derived CDR1/2 loops, and QC warnings
• exercise an upstream checkpoint on a current PyTorch/Lightning CPU stack in dedicated CI
• document why small CALIS parameters stay packaged and why DTU identity-bound registrations cannot be automated by --accept-license

The released MixTCRpred network consumes paired CDR3s and V-derived CDR1/2. J genes are accepted, corrected, and reported for schema completeness, but they are not network inputs.

Closes #230.
Related: #274.
Upstream runtime report: GfellerLab/MixTCRpred#10

Verification

• ./lint.sh
• ./test.sh — 734 passed, 56 skipped, 2 xfailed
• live current-runtime regression — 13 passed, including the real bundled checkpoint
• clean managed fetch, optional Zenodo download, checksum/provenance manifest, catalog listing, and the official 1,003-row example reproduced locally