db/state: invalidate BranchCache entries on Flush to fix stale trie reads

[erigon-copilot]

Summary

• Fixes wrong trie root regression in stage-exec-test (from-0, parallel) at block 263641 and (from-0, serial) at block 513814
• Root cause: SharedDomains.Flush() did not invalidate the BranchCache, so the integration-path from-0 loop (Flush + ClearRam + tx.Commit()) left stale commitment branches in the cache across batch boundaries
• Fix: Flush() now uses flushMemWithCallback to invalidate every commitment-domain key it writes, forcing subsequent reads to fall through to MDBX

Root cause

The BranchCache introduced in #21380 is checked before MDBX in GetLatest for CommitmentDomain. Commit() updates the cache after a successful commit, but Flush() — used by the integration stage_exec from-0 loop — left the cache untouched. When the same commitment branch was read (cached), updated, flushed, and then read again in the next batch, GetLatest returned the stale cached value instead of the freshly committed DB value.

Test plan

• TestBranchCacheFlushInvalidatesStaleEntries — new test that reproduces the exact stale-cache scenario; confirmed FAILS without the fix, PASSES with it
• TestBranchCacheCommitRefreshesAfterReadThrough — existing test, still passes
• TestFromZero_GenesisAllocPreservedAfterResetReExec — passes (serial + parallel)
• All db/state/execctx tests pass with -count=3 -race
• All execution/commitment tests pass with -count=3 -race
• make lint clean
• make erigon integration builds clean