Cache packer in delta writers (resolves TODO)

Addresses the TODO at DeltaBinaryPackingValuesWriterForLong.java:119:
"should this cache the packer?". Adds a per-instance BytePacker[] /
BytePackerForLong[] cache keyed by bit width to avoid a factory
dispatch + array load on each miniblock flush. Symmetric to the reader
cache added in the previous commit (9c14a33).

Results on IntEncodingBenchmark.encodeDelta (5 iter x 1 fork):
  SEQUENTIAL:       73.05M -> 73.67M  (+0.8%, within noise)
  RANDOM:           50.99M -> 51.51M  (+1.0%, within noise)
  LOW_CARDINALITY:  61.15M -> 61.37M  (+0.4%, within noise)
  HIGH_CARDINALITY: 51.30M -> 50.79M  (-1.0%, within noise)

Kept as code-quality improvement: resolves the explicit TODO, mirrors
the reader structure, and the long writer has a 65-entry bit-width
space vs 33 for int (more factory dispatch cost avoided).

All 573 parquet-column tests pass (including 100K-round random delta
tests).

Author: iemejia

parquet-column/src/main/java/org/apache/parquet/column/values/delta/DeltaBinaryPackingValuesWriterForInteger.java

@@ -60,6 +60,12 @@ public class DeltaBinaryPackingValuesWriterForInteger extends DeltaBinaryPacking
    */
   private int minDeltaInCurrentBlock = Integer.MAX_VALUE;
 
+  /**
+   * Cache of BytePacker instances indexed by bit width [0, 32].
+   * Avoids a factory dispatch + array load per miniblock flush.
+   */
+  private final BytePacker[] packerCache = new BytePacker[MAX_BITWIDTH + 1];
+
   public DeltaBinaryPackingValuesWriterForInteger(int slabSize, int pageSize, ByteBufferAllocator allocator) {
     this(DEFAULT_NUM_BLOCK_VALUES, DEFAULT_NUM_MINIBLOCKS, slabSize, pageSize, allocator);
   }
@@ -116,7 +122,11 @@ private void flushBlockBuffer() {
     for (int i = 0; i < miniBlocksToFlush; i++) {
       // writing i th miniblock
       int currentBitWidth = bitWidths[i];
-      BytePacker packer = Packer.LITTLE_ENDIAN.newBytePacker(currentBitWidth);
+      BytePacker packer = packerCache[currentBitWidth];
+      if (packer == null) {
+        packer = Packer.LITTLE_ENDIAN.newBytePacker(currentBitWidth);
+        packerCache[currentBitWidth] = packer;
+      }
       int miniBlockStart = i * config.miniBlockSizeInValues;
       // Mini blocks are always flushed as full 32-value groups in the current format.
       // Use the packer's 32-value entry point to avoid four pack8Values calls per miniblock.

parquet-column/src/main/java/org/apache/parquet/column/values/delta/DeltaBinaryPackingValuesWriterForLong.java

@@ -60,6 +60,12 @@ public class DeltaBinaryPackingValuesWriterForLong extends DeltaBinaryPackingVal
    */
   private long minDeltaInCurrentBlock = Long.MAX_VALUE;
 
+  /**
+   * Cache of BytePackerForLong instances indexed by bit width [0, 64].
+   * Avoids a factory dispatch + array load per miniblock flush.
+   */
+  private final BytePackerForLong[] packerCache = new BytePackerForLong[MAX_BITWIDTH + 1];
+
   public DeltaBinaryPackingValuesWriterForLong(int slabSize, int pageSize, ByteBufferAllocator allocator) {
     this(DEFAULT_NUM_BLOCK_VALUES, DEFAULT_NUM_MINIBLOCKS, slabSize, pageSize, allocator);
   }
@@ -116,8 +122,11 @@ private void flushBlockBuffer() {
     for (int i = 0; i < miniBlocksToFlush; i++) {
       // writing i th miniblock
       int currentBitWidth = bitWidths[i];
-      // TODO: should this cache the packer?
-      BytePackerForLong packer = Packer.LITTLE_ENDIAN.newBytePackerForLong(currentBitWidth);
+      BytePackerForLong packer = packerCache[currentBitWidth];
+      if (packer == null) {
+        packer = Packer.LITTLE_ENDIAN.newBytePackerForLong(currentBitWidth);
+        packerCache[currentBitWidth] = packer;
+      }
       int miniBlockStart = i * config.miniBlockSizeInValues;
       // Mini blocks are always flushed as full 32-value groups in the current format.
       // Use the packer's 32-value entry point to avoid four pack8Values calls per miniblock.