refactor(train): share per-micro anchor preparation between train stacks

[chi030303]

Summary

TrainStack and UnifiedModelTrainStack both froze π_old by slicing planned micros, calling algorithm.prepare_segment, rejecting a missing anchor_fields tensor, then concatenating in plan order. That stitch now lives in one helper next to _validate_anchor_contract. Each stack still builds its own micro_slices (Plan vs _optimizer_step_slices). old_logp_source=rollout, packed concat order, and the single-micro / non-recompute fast paths are unchanged.

Related Issue

Closes #439

Test Plan

• python -m py_compile unirl/train/stack/base.py unirl/train/unified_model_stack.py — pass
• Uncommitted packed-varlen smoke (not added under tests/): multiple micros with non-uniform packed lengths, token-budget planner concat order, num_updates_per_batch>1, missing declared field, old_logp_source=rollout whole-shard, TrainStack and UnifiedModelTrainStack — 6/6 passed
• Live train: Not run; reason: refactor-only, no recipe or algorithm change

Compatibility / Risk

No config, checkpoint, data format, or public API changes. Unified missing-field errors now say "a micro produced None" instead of "a micro-slice produced None"; the stack prefix is unchanged.

Reviewer Notes

Proposed on #439 before opening this PR. No overlapping open PR touches this stitch (#476 edits eval_track only). AI-assisted implementation; I reviewed the diff. Follow-up commit cd2cbc2 is ruff-format only (error string on one line).

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.