feat(monitor): unify explorer/trainer wandb logging into one run

[MengsD]

When monitor.monitor_type is wandb and monitor.shared_run is True, explorer and trainer log their metrics to the same run; metrics are labelled by {role}/step.

Description

[Please describe the background, purpose, changes made, and how to test this PR]

Background

Closes #556

Why this design

A wandb run has a single global _step — a per-run counter assigned in arrival order. With two roles writing one shared run, their rows interleave, so each role lands on a gapped _step (e.g. explorer 0,2,4…, trainer 1,3,5…) and can't show a clean 1..N. The only fix is to give each metric its own x-axis via define_metric(key, step_metric="{role}/step").

Three constraints then shaped the implementation:

• No glob. wandb's define_metric glob doesn't cross / (issue #9549), and a bare * can't map the two roles to different axes — so bindings must be declared per full key, and keys appear dynamically during the run.
• Multi-process. explorer / trainer / launcher are separate processes, so writing one run requires wandb shared mode (one primary = the launcher, secondaries = the actors).
• Config is overwritten, not merged. In shared mode each writer's config sync replaces the run's metric defs (last writer wins). So a single declarer is unstable: if only the primary declares, the secondaries clobber it (the live axis falls back to the gapped _step); if each secondary declares only its own role, the two secondaries clobber each other.

Approach: make every writer carry the full binding set, so no sync can drop anything.

• Secondaries declare the full set on each flush (their own keys + all roles' keys read from shared binding files) → axis is correct during the run.
• The launcher (primary) declares the full set once before finalize → axis is correct after the run (the primary's config wins at finalize).

Changes made

• Add monitor.shared_run (opt-in, default False) and an internal monitor.run_id to MonitorConfig.
• The launcher creates one primary shared run and injects run_id; explorer/trainer join it as secondary writers (wandb shared mode).
• Each role's metrics are prefixed {role}/ and bound (via define_metric) to their own {role}/step axis, so explorer and trainer each show a clean 1..N x-axis within the same run — both during the run and after it finishes.
• The config validator downgrades shared_run to False for backends that don't support it (only wandb does); shared_run=False keeps the original behavior (separate runs) byte-for-byte.
• Note: the per-role step bindings are exchanged via files under checkpoint_job_dir/monitor/, so multi-node runs require that dir on shared storage (e.g., NAS).

How to test

1. In an example config (e.g. examples/gsm8k-quick.yaml) set:
   monitor:
   monitor_type: wandb
   shared_run: true
2. Run a both job: trinity run --config examples/gsm8k-quick.yaml.
3. Open the wandb run and confirm: explorer and trainer metrics live in one run; each role's panels use its own {role}/step axis showing 1..N (verified both mid-run and after finish).
4. With shared_run: false (default) or a non-wandb backend, behavior is unchanged (two separate runs).

Checklist

Please check the following items before code is ready to be reviewed.

• Code has passed all tests
• Docstrings have been added/updated in Google Style
• Documentation has been updated
• Code is ready for review

[MengsD]

/unittest-diff

[github-actions]

unittest: Run #1780

Tests 📝	Passed ✅	Failed ❌	Skipped ⏭️	Pending ⏳	Other ❓	Flaky 🍂	Duration ⏱️
203	201	0	2	0	0	0	39m 46s

🎉 All tests passed!

Github Test Reporter by CTRF 💚