Cardano Updates

Add certificate-based filtering to the watch module, enabling clients
to monitor transactions containing specific certificate types (stake
registration, delegation, pool registration, DRep, etc).

Includes tests using real mainnet delegation tx CBOR data.

The predicate was already derived from start_shard >= total_shards but
its name described the consequence in finalize ("replay no-op") rather
than the state ("shard work is complete"). With the state framing, the
same predicate is meaningful in initialize too: assign the fields
unconditionally from progress, then check is_complete() for the
short-circuit. Removes the special-cased branch that previously set
start_shard = p.total inline.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

- Fix EpochWrapUpV2 docstring: variant index is 42, not 40 (counting
  past EWrapProgress=39, EStartProgress=40, RupdProgress=41).
- Add TODO(wal-compat) block above EpochWrapUpV2 mirroring the V1
  block, so the V3 cutover has the same documented removal trigger
  ("no on-disk WAL row references variant index 42").
- Document is_replay_noop() precondition: must be called after
  initialize(); a freshly-new()'d unit would spuriously report true.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

EpochState had no flag distinguishing "EWRAP not started" from "EWRAP
done, ESTART pending finalize" — both showed `ewrap_progress = None`.
A crash between EWRAP finalize and ESTART finalize left the cursor
behind the boundary block (cursor only advances at ESTART finalize),
so the next live block re-triggered EWRAP for the same boundary. The
replay ran visitors over partially-rotated accounts and panicked at
`PoolDelegatorRetire::apply`'s `unreachable!`.

Add `EpochWrapUpV3` that no longer clears `ewrap_progress` on apply —
the field stays at `Some(total, total)` after EWRAP finalize and is
cleared only by `EpochTransitionV2` at ESTART finalize. That single-
writer invariant makes `committed == total` the unambiguous "EWRAP
done, ESTART pending" marker. `EwrapWorkUnit::initialize` reads it,
sets `start_shard = total_shards` to skip the shard loop, and a new
`is_replay_noop()` accessor lets `finalize` no-op too. The next
boundary trigger falls through to ESTART, which resumes from
`estart_progress.committed` via the existing per-shard path.

V2 stays around `#[deprecated]` for WAL replay of historical rows,
mirroring the V1 → V2 cutover pattern.

Follow-up TODOs left at `PoolDelegatorRetire::apply` and
`DRepDelegatorDrop::apply` for the deeper hardening (read via
`snapshot_at(self.epoch)`; guard schedule on `entity.epoch() ==
Some(self.epoch)`) — the recovery short-circuit removes the known
trigger but the visitor/delta read divergence is still fragile.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>