feat(loggers): add SwanLab logger as a wandb alternative (--swanlab_args)

[Luodian]

Add SwanLab logger as a Weights & Biases alternative

Adds an optional SwanLab results logger for evaluation runs, wired symmetrically to the existing wandb logger and driven by a new --swanlab_args flag. SwanLab is a lightweight, self-hostable experiment tracker many users prefer as a W&B alternative.

Usage

python3 -m lmms_eval \
  --model llava --model_args pretrained=... \
  --tasks mme \
  --swanlab_args project=lmms-eval,exp_name=my-run,mode=cloud \
  --swanlab_log_samples

Install the extra: pip install lmms-eval[swanlab] (or pip install swanlab).

Authenticate with the standard SwanLab mechanism — set SWANLAB_API_KEY (and optionally SWANLAB_HOST for a self-hosted instance). The default project is lmms-eval, overridable via the args string (project=, exp_name=, mode=, and any other swanlab.init kwarg).

What it does

• SwanLabLogger (lmms_eval/loggers/swanlab_logger.py) mirrors WandbLogger's interface: post_init(results), log_eval_result(), log_eval_samples(samples), finish().
• log_eval_result() logs task configs to the run config and the flattened per-task numeric metrics as SwanLab scalars.
• log_eval_samples() (enabled with --swanlab_log_samples) logs a per-task sample count plus a capped JSON preview via swanlab.Text.
• Wiring in lmms_eval/__main__.py is parallel to wandb: init when --swanlab_args is non-empty, log on the main process after each eval (best-effort, wrapped in try/except), finish at the end.

Notes

• The module imports fine without swanlab installed; swanlab is imported lazily and raises a clear ImportError (pip install swanlab) only if the logger is actually used without it.
• SwanLab (0.7.x) has no Table/Artifact type, so wandb's results-as-table and results-as-JSON-artifact steps are not replicated — the scalar metrics cover the leaderboard use-case.
• No changes to the existing wandb code paths.

Validation

• py_compile on all changed files; module imports cleanly without swanlab installed (lazy import, clear ImportError on use)
• Functional smoke test with real swanlab 0.7.15 in mode=offline: init → post_init → log_eval_result → log_eval_samples → finish all pass on representative results/samples dicts
• ruff clean on the new module; black + isort (pre-commit) pass

Review pass (2026-07-06)

Two self-review commits:

• Best-effort init: SwanLabLogger(...) construction in __main__.py is now wrapped like the log calls — a bad SWANLAB_API_KEY/SWANLAB_HOST logs a warning and the eval continues without swanlab, instead of aborting a run at startup (swanlab.login() fires during init whenever SWANLAB_API_KEY is set, even for offline runs).
• swanlab >= 0.8 support: on 0.8.x swanlab.get_run() raises when no run is active (0.7.x returned None), which would have sent every fresh run into the failure path. Both behaviors are handled; verified against real swanlab 0.8.4 (fresh construction reaches swanlab.init exactly once) and 0.7.x semantics are preserved.