Fix the DB-level unwind path so first-time-write reversals leave a proper
empty-value tombstone in the domain values table, rather than deleting the
current-step entry and silently falling back to files (which have no
concept of deletions and therefore return stale pre-unwind data).

`DomainEntryDiff.Value` carries three distinct shapes that must be handled
differently during unwind (per the comment in db/kv/helpers.go):

- nil         — "different step": prev value lives at another step, skip
                  restore (only produced by legacy V0 changesets where
                  valueLen==0 deserializes as nil)
- []byte{}    — "no previous value": key was absent before this step; write
                  an empty tombstone at the unwind step so subsequent reads
                  see the key as absent again
- non-empty   — restore the actual previous value

The old guard `if len(value) > 0` conflated `nil` and `[]byte{}` and
skipped both, leaving no tombstone after unwinding a first-time write.
Corrected to `if value != nil`, matching the diff's documented contract.

The written tombstone lands at max(step, currentFilesEndStep), which by
construction is beyond the frozen file range, so the existing step-age
guard in getLatestFromDb already accepts it without needing to change the
read path.

This is a deliberately narrow re-land of #20483 (reverted by #20509). The
PR's other two pieces — the `len(v)==0` empty-authority bypass in
getLatestFromDb and the LargeValues cross-check in getLatest — are NOT
included here. Regressions surfaced in kurtosis/assertoor_glamsterdam
(which runs with --experimental.bal and the storagerefundtx spammer,
generating a deluge of tombstones) suggest the empty-authority bypass
has downstream interactions with the BAL / post-Amsterdam paths that
need separate narrowing. The interrupted-prune LargeValues case is the
original regression suspect from #20509 and remains out-of-scope.

Tests in db/state/domain_test.go:

- TestDomain_UnwindRestoresDeletionMarker (DupSort + LargeValues) — write
  a key, delete it, re-write within the same step, build files, then unwind
  the re-write. Fails on pre-fix code (getLatest returns the stale post-
  unwind value from the frozen file); passes with this fix. Directly
  exercises the write-side bug.

- TestDomain_DeletedKeyNotResurrectedByFiles (DupSort + LargeValues) —
  retained as a forward regression guard for the read-side contract,
  even though it passes on both pre-fix and post-fix code on current main
  (file-build + prune semantics have evolved since #20483 such that the
  original scenario no longer triggers the read-side bug on its own).

Tested: go test -short ./db/state/... all pass; make lint 0 issues;
make erigon builds.

Remove 3 test helpers/tests that depend on main-only APIs (beginForTests,
_testBuildAccessorHook). Replace beginForTests() with BeginFilesRo() in
the two tests that are part of this cherry-pick.