Bug fixes: composite PK, nl_to_sql sandbox, FK check, SQL injection, storage correctness

Critical fixes:
- Composite PK: execInsert + validateConstraints use all PK columns
- nl_to_sql: non-SELECT SQL no longer executed directly during validation
- FK check: removed O(N) scanMemTable fallback, uses db.get() with SSTables
- exprToSql: nkIdent wrapped in quotes to prevent SQL injection
- restoreSchema: try/except around tokenize/parse for crash resilience
- recovery.nim: lastTxnId tracking + putUnsafe/deleteUnsafe without WAL
- SCRAM: verifyClientProof length check + DefaultIterationCount restored

Storage fixes:
- lsm.nim: SSTable sort order fixed (ascending), close() flushes all memtables
- compaction.nim: tombstones preserved during compaction
- wal.nim: header written for empty existing files, readEntries checks magic
- btree.nim: B+ tree leaf split keeps boundary key
- bloom.nim: deserialize raises on short data
- mmap.nim: bounds checks for adviseWillNeed/DontNeed + posix.close()

Protocol fixes:
- zerocopy.nim: readString bounds check
- wire.nim: deserializeValue 32-bit underflow check
- auth.nim: JWT JSON escaping for claims
- server.nim: readUint32BE bounds check + specific exception handling
- raft.nim: readData checks + specific exception handling

Tests:
- Added Composite Primary Key test suite (4 tests)

Build: 0 warnings, 0 errors

src/barabadb/storage/compaction.nim

@@ -98,11 +98,10 @@ proc compact*(cs: CompactionStrategy, level: int): CompactionResult =
       merged.add(entry)
       lastKey = entry.key
 
-  # Filter out tombstones (deleted entries)
+  # Keep tombstones to prevent deleted keys from resurrecting in lower levels
   var final: seq[Entry] = @[]
   for entry in merged:
-    if not entry.deleted:
-      final.add(entry)
+    final.add(entry)
 
   # Write merged SSTable
   let outputPath = cs.dataDir / "sstables" / ("level_" & $level & "_" & $tables[0].createdAt & ".sst")