fix(dashboard): isolate telemetry per training run

[Rockdu]

What

• Dashboard telemetry now lands in a per-launch subdirectory <workspace>/run_<UTC timestamp>/ instead of directly in the workspace root.
• python -m miles.dashboard.viewer --workspace <dir> resolves to the newest run_* subdirectory; pass a specific run_* directory to look at an older run.
• No flag added or removed. --miles-dashboard-workspace keeps its meaning (it is the base directory) and is never rewritten behind the user's back.

Why

DashboardStore partitions each stream by UTC hour and appends (phases/20260728_22.jsonl, ...), and meta.json is written once per collector. Relaunching training with the same --miles-dashboard-workspace therefore appended the new run's phases, trajectories and GPU samples into the previous run's files and overwrote meta.json, so the viewer drew two runs as a single timeline — the reported symptom.

Two design points worth a look during review:

• The run directory is derived on the driver only. Instead of rewriting args.miles_dashboard_workspace (that would silently change a user-supplied path, and every worker would inherit the mutated value), the collector owns the run directory and exposes it through DashboardCollector.workspace(). The rank-0 GpuUtilSampler — the only worker that writes telemetry files itself — asks the collector for it; phase and trajectory events already flow through the collector, so nothing else needs the path.
• Run directory names are second-granular and sort lexicographically, which is what resolve_run_dir relies on to pick "newest".

For reference, miles core does not handle this inside the dashboard either: its telemetry directory is {dump_details}/dashboard with --dump-details asserted as required, and per-run uniqueness comes from the launcher (--dump-details {output_dir}/{run_id}/dump_details, run_id = UTC timestamp plus a random suffix). miles-diffusion has no such layer — no example or script passes --dump-details — and its viewer reads only the JSONL streams, so deriving the workspace from --dump-details would force heavyweight per-rollout .pt dumping for no benefit. This PR is the same idea one level down: one timestamped directory per launch, owned by the dashboard instead of the launcher.

Known limits, deliberately not addressed here:

• Two launches whose start timestamps land in the same UTC second compute the same run directory and merge. A sequential relaunch cannot realistically hit this; two drivers started concurrently can. There is no arbitration today: the driver only derives the path, the directory itself is created lazily by DashboardStore.write_meta with exist_ok=True. Guarding it means having the driver create the directory with exist_ok=False and fall back to a -N suffix.
• Telemetry written into the workspace root before this change stays there; once any run_* directory exists the viewer resolves to that instead, so point --workspace at the old directory itself to read it.

Files

• miles/dashboard/store.py — run_dir() (launch timestamp to directory) and resolve_run_dir() (workspace to newest run).
• miles/dashboard/backend.py — the driver points the collector config at this launch's run directory.
• miles/dashboard/collector.py — workspace() accessor so a worker can locate the run directory.
• miles/dashboard/hooks.py — rank-0 GPU sampler writes into the collector's run directory.
• miles/dashboard/viewer.py — --workspace resolves to the newest run.
• miles/utils/arguments.py — help text for --miles-dashboard-workspace.
• tests/fast/dashboard/test_run_isolation.py — new: a relaunch does not merge metrics, newest-run resolution (including a workspace that holds no runs), and init_dashboard leaving the workspace arg untouched.
• tests/fast/dashboard/test_async_collector.py — drop a stale monkeypatch.setattr(backend, "_is_primary", ...); the attribute no longer exists, so that test was erroring on main as well.

Validation

• pytest tests/fast/dashboard -q → 11 passed (8 existing, 3 new).
• Manual viewer check: created two run directories under one workspace; --workspace <base> rendered only the newest run's phase span, --workspace <base>/run_<older> rendered the older one.
• Not run locally: the full tests/fast suite and train_diffusion.py --help — this checkout has no ray/torch, so collection fails at import for both. Both need CI or a GPU box.

Checklist

• pre-commit run --all-files
• Added/updated tests for new behaviour
• pytest -x is green
• If launch flags changed, python3 train.py --help still parses — not run (needs ray); no flag was added or removed, only the help text of an existing one
• If a public flag was added, it appears in the CLI reference docs — n/a, no new flag
• If an example was added, it has a real walkthrough — n/a, no example