[Proto] Add a Fit step and FitCohort

[jrapin]

Fit: one artifact over a cohort of items

Fit derives an artifact from many items, then transforms each item with it — normalization stats, a PCA basis, a trained model.

class Normalize(steps.Fit):

    def _fit(self, values):    # the cohort
        stacked = np.concatenate(list(values))
        return stacked.mean(0), stacked.std(0)

    def _run(self, value):     # one item
        mean, std = self.fitted
        return (value - mean) / std

norm = Normalize(infra={"backend": "Cached", "folder": cache})
norm.run_many(steps.FitCohort(train))  # fits on these, then transforms them
norm.run_many(test)                    # same artifact, novel items

• Only a FitCohort is fitted on; anything else transforms with the artifact already fitted.
• The cohort's fingerprint — or its name, Normalize(cohort="train") — scopes the artifact and every downstream cache. A name is recoverable from the config alone; another cohort needs another config (clone()).
• The fit runs before any backend splits the batch, cached under infra. The upstream is read twice, so give an expensive one its own infra.