…lse positives

TestDeleteRecreateSlotsAcrossManyBlocks under GOMAXPROCS=2 -race flaked
~3-5% with EXEC3_PARALLEL=true, in two modes: a +1602 gas mismatch (a
CREATE2 collision burned inner gas on a stale CodeHash read) and a
runaway validator-invalid retry loop.

Changes:

  execution/state/intra_block_state.go
    - getStateObject initialised accountVersion to sdb.Version() (the
      worker's own incarnation). refreshVersionedAccount's sub-read
      updates are gated on `bversion >= readVersion`, so every prior-tx
      write at a lower version was silently discarded, leaving
      accountSource pinned to StorageRead with a worker-private version
      vm never wrote. Switched to UnknownVersion so sub-reads promote.
    - Removed the synthetic AddressPath read recorded by accountRead: it
      borrowed the version that flowed through refreshVersionedAccount
      (often a sub-read's), tripping the validator's path==AddressPath
      checks against entries vm never wrote. Sub-reads are recorded by
      versionedRead directly.

  execution/state/versionedio.go
    - Exclude AddressPath from the readStorage==nil MVReadResultNone probe
      recording. getStateObject probes vm for AddressPath then falls back
      to the stateReader (the account does exist); recording the nil
      probe as "we read nothing" made the validator's path==Address
      BalancePath-presence check invalidate any pre-existing account a
      prior tx wrote BalancePath for.

  execution/state/versionmap.go
    - validateRead split into validateReadImpl with a `recursive` flag.
      Two false-positive paths in the cross-validate recursion produce
      spurious Invalid when invoked as a synthetic probe (readVal nil,
      source borrowed): the path==AddressPath BalancePath-presence check,
      and the source!=MapRead+nil-readVal tiebreaker. Both gated on
      !recursive.
    - Added an SD-staleness cross-check in the MVReadResultDone branch:
      a later tx self-destructing without revival means the read predates
      the destruct and the serial path returns zero via the SD-zero
      short-circuit; checkVersion alone misses it because SD doesn't
      write the read's own path. revivalLimit = txIndex-1 excludes the
      validating tx's own writes (validation runs post-flush).

  execution/stagedsync/exec3_parallel.go, exec3_status.go
    - On ErrDependency with no effective blocker (named blocker already
      complete) AND on validator-invalid, push the tx to a new deferred[]
      queue instead of immediate pending. scheduleExecution's drain
      predicate gates retry on (predecessor validated) AND (no worker at
      a lower index in flight) — lower-indexed workers' flushes land at
      indices visible to N's reads via vm.Read's floor(N-1). Non-deferred
      txs keep dispatching speculatively. A safety net force-drains when
      pending is empty and no workers are in flight.
    - tooManyRetries helper + cap on validator-invalid retries, mirroring
      the worker-error path, so a genuine loop can't spin forever.

Stress: 200/200 on TestDeleteRecreateSlotsAcrossManyBlocks (was
95-97/100 baseline); no regression on the EEST blocktest suite under
the 12-worker parallel runner.

…r-validator

# Conflicts:
#	execution/state/versionmap.go

Covers the SD-staleness cross-check in validateReadImpl: a version-
consistent MapRead is invalidated when a later tx self-destructed the
account with no revival, and stays valid when a revival write follows
the destruct. Also documents the valuesEqual tiebreaker branch.