Add xToken-aligned cross-tokenizer OPD

[artem-osmosis]

Summary

• port Slime's OPD feature parity into Miles, including sampled-token and top-k training data/loss paths, validation, metrics, and eval reward dispatch
• add xToken-style cross-tokenizer alignment with independent student/teacher chat templates, strict/retried teacher scoring, zero-signal masking, and alignment coverage telemetry
• preserve OPD fields through multi-turn, CP, and multimodal preprocessing, including response-side media expansion safeguards
• add the single-node 8×H200 Llama-3.2-3B / Nemotron-3-Super-120B launcher with scoped Ray/teacher cleanup, safe path guards, and a shared sglang-miles API preflight
• document the new modes and add focused unit coverage

Validation

• 8×H200 end-to-end run from checkpoint iteration 79 using sglang-miles-v0.5.13-final at f8cfad35bdfeeed09b259857e738bf833aeb70e0
• native /begin_weight_update and /end_weight_update transactions returned 200; zero begin 404s and zero legacy /post_process_weights calls
• 10 optimizer steps (80–89), 10 alignment batches, zero alignment fallbacks
• prescribed reverse-KL trend check:
  • start: 0.3680919334
  • end: 0.2980055362
  • delta: -0.0700863972
  • first-half mean: 0.3562976375
  • second-half mean: 0.3052670628
  • slope: -0.0089484361 / step
  • start-in-range: true
• W&B: https://wandb.ai/yatsenkoa/miles-opd/runs/7ow8s81i
• relevant suite: 149 passed, 1 skipped
• CI registration suite: 60 passed
• ruff check, git diff --check, and launcher bash -n pass
• incompatible SGLang source, overlong Ray socket path, unsafe/pre-existing Ray temp directory, and BSHD+CP multimodal cases fail closed before unsafe work