…lesInBackground

Port of #20445 to release/3.4.

Fixes two related bugs causing gas mismatches during execution (#20169).

Bug 1: Collation/pruning race — BuildFilesInBackground opened a new
read-transaction per domain/index, so execution could commit step S+1
data between reads. The collated file for step S then contained wrong
values; after pruning, GetLatest returned stale file values.
Fix: restructure buildFiles into two phases — Phase 1 collates all
domains and indices using a single MDBX read-transaction; Phase 2
builds files in parallel from the collations (no DB access).
Also adds a committedTxNum guard: don't collate step S until
ComputeCommitment has confirmed all data through end of step S is flushed.

Bug 2: Unwind entry visibility — after a reorg, restored entries were
tagged with the natural step, which may already be covered by files.
getLatestFromDb discarded those entries and fell through to stale file values.
Fix: pass Aggregator.EndTxNumMinimax() into unwind and tag restored
entries with max(naturalStep, currentFilesEndStep).

Also: export DecodeTxBlockNums; fix minUnwindale typo; short-value guard.

Three unit tests that each FAIL against the current implementation,
proving the PR has correctness issues:

1. TestStepFullyCommitted_ZeroIsGenesisNotSentinel
   stepFullyCommitted(committedTxNum=0, ...) returns true (bypass) for any
   input, conflating "genesis committed at txNum=0" with "no commitment yet".
   When genesis writes committedTxNum=0, ALL steps are bypassed — the guard
   is completely ineffective for the step that genesis belongs to.

2. TestDomain_CollationRaceNotTestedByIsolationTest
   TestDomain_CollationIsolatedFromLaterSteps tests cross-step filtering
   (trivially correct from bytes.Equal check), NOT the actual partial-commit
   race described in #20169. The real race: step S data in two separate DB
   commits, collation opening its read tx between them → second batch invisible
   → data permanently lost after pruning. Single-tx collation alone doesn't
   prevent this when committedTxNum=0 bypass allows premature collation.

3. TestDomain_UnwindOverBumpedStepMasksNewWrites
   AggregatorRoTx.Unwind uses at.a.EndTxNumMinimax()/StepSize() (current
   global minimax) rather than the per-domain snapshot FirstStepNotInFiles().
   If BuildFilesInBackground filed a new step AFTER the AggregatorRoTx was
   opened, currentFilesEndStep is over-bumped (e.g. 2 instead of 1). The
   restored unwind entry lands at step 2. Post-reorg writes at step 1 do NOT
   replace the step-2 entry (different step byte encoding). getLatestFromDb
   then returns the stale step-2 restored value instead of the correct
   post-reorg step-1 value.