feat(datagen): reshape checkpoint delta-log to schema v2

[YangjunZ]

What & why

The datagen checkpoint delta-log had drifted from the authoritative spec. This
PR brings the lance-context store in line with schema v2, and — separately
from the code — makes the change reviewable by documenting the model and adding
the tests the reshape was missing.

The storage mechanism is unchanged: still a single append-only log.lance,
MemWAL sharding, deterministic event_id, blob offload. Only the event
vocabulary and the folded model change.

Changes

Event vocabulary (6 → 7 events)

• Add STEP_STARTED — a structural driver-frame marker (Sequence/Loop
  only). started \ completed is exactly the frame open at crash time, which is
  what resume re-enters.

Status is a first-class column

• Replace the terminal column (only completed/filtered) with a status
  column carrying four values: running / completed / filtered / failed.
  running is now stored on ITEM_CREATED rather than derived.
• Two read lenses: lifecycle (fold; ignores FAILED) vs failure (reads
  FAILED rows). A failed step leaves the item running under lifecycle and
  surfaces only through the failure lens.

Structured provenance

• Replace opaque step_instance_id + iteration with step_kind /
  enclosing_step / selector_step. A resume can now rebuild step
  coordinates instead of trusting a stored cursor blob.
• DatagenItemId is a structured, materialized path (5/solve_twice:0); ids
  compose purely on the client with no store round-trip.
• New DatagenStepKind { Root, Leaf, Sequence, Loop, MapReduce, Branch, SubPipeline, Conditional, Router }. Only drivers emit STEP_STARTED; fan-out
  kinds emit only their reduce STEP_COMPLETED; selectors emit no row and the
  chosen child records selector_step.

Fold semantics

• fold_datagen_events returns Option — None (NeverStarted) when there is
  no ITEM_CREATED, which is the fresh-vs-resume fork.
• FIELD_SET last-writer-wins; FIELD_APPEND accumulates; mixing the two on one
  field is rejected. DATAGEN_SCHEMA_VERSION = 2.

Docs

• specs/datagen-checkpoint-schema.md — rewritten for v2: the seven events,
  the status column, structured item id + step provenance, the two read lenses,
  and updated write/recovery invariants.
• docs/design/using-datagen-store.md (new) — a client-facing walkthrough:
  open → compose ids → write/checkpoint → finish → fold/resume → read tree /
  classify roots / failures / lazy blobs, all with runnable Rust snippets, plus a
  kind→what-you-emit table.

Tests

Fold-layer (datagen.rs) and store-layer (datagen_store.rs) coverage grew
from 12 → 19:

• step_kind / status parse round-trips + is_driver
• TERMINAL → filtered status
• selector_step folded onto the chosen child's position
• fan-out sub-item folds with denormalized parent/root lineage
• FIELD_SET / FIELD_APPEND mixing is rejected
• resume open-frame = started \ completed
• store: whole fan-out tree read by root + root_item_statuses classification +
  NeverStarted sibling

test result: ok. 19 passed; 0 failed

Compatibility

Schema v2 is not backward-compatible with v1 logs (the terminal column and
step_instance_id/iteration provenance are gone). This targets new
experiments; there are no v1 logs in production to migrate.