Add b2luigi dependency and workflow prerequisites (gitea #83)

Groundwork for the b2luigi pipeline orchestration in gitea #83, split out
so the workflow package itself lands as a self-contained change:

- new `workflow` optional-dependency extra (b2luigi, which pulls luigi +
  tenacity), included in `dev`.
- deterministic rollout/predict sidecar path: with an explicit `--out`, the
  YAML goes to `out.with_suffix(".yaml")` instead of a uuid-named file under
  the checkpoint directory, so a workflow task can declare it as a target.
  The uuid behaviour is kept for the no-`--out` case, leaving ad-hoc runs and
  the /ceph predictions convention untouched.
- epoch-aware shuffle seeding in StreamingStepsDataset (`seed` +
  `set_epoch`, the DistributedSampler convention). Shuffling previously drew
  from the global numpy state, which `run_train_job` reseeds from
  `train.seed` at process start — so a one-epoch-per-job chain would have
  replayed the same batch order every epoch. Seeding from
  `(seed, epoch, worker_id)` makes epoch k's order identical whether it runs
  inside one long `giant train` or as its own resumed job. The val-split
  seed is untouched.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

pyproject.toml

@@ -28,7 +28,7 @@ dev = [
     "ty>=0.0.50,<0.1",
     "bump-my-version>=1.2,<2",
     "git-cliff>=2,<3",
-    "giant[convert,analysis,geometry,wandb]",
+    "giant[convert,analysis,geometry,wandb,workflow]",
 ]
 geometry = [
     "scikit-learn>=1.4,<2",
@@ -49,6 +49,11 @@ analysis = [
     # `giant analyze render` step imports it; compute workers never do.
     "plotstyle>=1.0.0",
 ]
+# b2luigi pulls luigi + tenacity; the only sanctioned way to chain a
+# multi-step pipeline (see giant/workflow/).
+workflow = [
+    "b2luigi>=1.0,<2",
+]
 
 [project.scripts]
 giant = "giant.cli:app"