v1.0.6: Support custom id field in input json

What's Changed

• Auto-detect CUDA architectures for layer norm kernel compilation by @longleo17 in #252
• Featurizer: zero-copy tensor creation + numpy vectorization by @longleo17 in #246
• MSA encoding: vectorized sequence parsing by @longleo17 in #247
• Dataset: vectorize pandas/numpy operations by @longleo17 in #248
• Attention precision fix + LDDT fused thresholds + loss caching by @longleo17 in #249
• Fused Triton dropout+residual-add kernel (v2, cueq fix) by @longleo17 in #256
• Fix/blackwell gpu compatibility by @giulioisac in #257
• chain reindexing by @giulioisac in #260
• fix return identity matrix in eye_mask when opposite is False by @mrzzmrzz in #261
• Remove invalid AdamW arg and handle None param_names by @mrzzmrzz in #262
• Support custom id field in input json

Full Changelog: v1.0.5...v1.0.6