#3+#4 (correctness/robustness): the sparse-vs-Patricia agreement counter lived
in static fields on FlatWorldStateScope. Every scope (read-only RPC, parallel
block processing, snapshot validation) updated and read the same counter, so
one scope's mismatch could reset another's confidence and one scope's wins
could promote another to authoritative without that scope validating its own
sparse trie. Moved to SparseAuthoritativeTracker held by FlatScopeProvider so
each provider owns its own confidence.

Also tied _sparseIsAuthoritative to SparseTrieSkipPatricia: the previous
condition (just 10 consecutive matches and !VerificationMode) flipped sparse
into the authoritative root/commit role while Patricia BulkSet still ran on
every block, paying for trie work whose result was thrown away. Now sparse
takes over only when the operator has explicitly opted out of Patricia
persistence via SparseTrieSkipPatricia (currently default-on for the EXPB
sparse-storage runs).

#5 (performance): storage commit walked every retained sparse storage trie
on every block. With cross-block PreservedSparseTrie reuse and no LFU pruning
yet, that's O(retained contracts) per commit and grows unboundedly. Expose
DirtyStorageAccountHashes (a lock-free enumeration of SparseRootComputer's
_storageChanges keys) and walk only the contracts whose storage actually
changed this block. CommitStorageTrie still skips clean subtrees internally
as a second filter, but now we don't even iterate cold contracts at the outer
level.

Test updates: FlatWorldStateScopeProviderTests registers SparseAuthoritativeTracker
in the Autofac container; FlatOverridableWorldScope and WriteBatchBenchmark
pass new instances at construction. All 432 State.Flat tests pass.