[feat] multiplex trainer token export and agg to prime monitor on orch side

[Jackmin801]

Note

Medium Risk
Changes distributed export layout, cross-rank STABLE coordination, and orchestrator filesystem metrics; low production impact but affects multi-run debugging and monitoring correctness.

Overview
Multi-run token exports now land under each run’s directory (output_dir/<run_id>/token_exports/step_<run_step>/) instead of only the trainer root, with run_id / run_step stamped on micro-batches in the packer and carried through transport and the data loader.

The trainer marks export steps complete via a distributed STABLE file after each step (mark_stable at end of forward/backward), and export records gain export_step and run_id. Non-exporting CP ranks still participate in stable marking.

The orchestrator reads stable token_exports JSONL and logs aggregated entropy and mismatch_kl (mean/max over loss-masked tokens) to the monitor, one stable step at a time. Docs and unit/integration tests cover run-local paths and metrics.

^{Reviewed by Cursor Bugbot for commit d6bb6bb. Bugbot is set up for automated code reviews on this repo. Configure here.}