refactor: optimize parquet library read/write performance

- Batch all binary encode/decode functions to operate on arrays instead
of single values, reducing pack/unpack call overhead
- Replace BinaryBufferReader's DataSize objects and Bytes allocations
with raw int tracking and string returns, eliminating object creation on
hot paths
- Skip Dremel shredding/assembly for flat non-nested columns, bypassing
unnecessary FlatValue object creation and definition level conversion
- Replace array_merge with array_push in ColumnChunkBuilders and batch
statistics tracking to avoid O(n²) array copying
- Memoize maxDefinitionsLevel/maxRepetitionsLevel on schema columns and
cache flatPath in local variables to cut repeated method calls

refactor: parquet writer from row-by-row to columnar shredding

refactor: replace serialize with pack on primivite values

refactor: optimize flat columns writing

refactor: make dremel shredder to precache shredding plans

Author: norberttech

shell.nix

@@ -48,6 +48,7 @@ pkgs.mkShell {
         pkgs.figlet
         pkgs.symfony-cli
         pkgs.act
+        pkgs.hyperfine
     ]
         ++ pkgs.lib.optional with-blackfire pkgs.blackfire
         ++ pkgs.lib.optionals with-wasm [