feat(PartitionedOutput): Add nested RowVector support in the serializer

[yingsu00]

No description provided.

[libianoss]

@xin-zhang2 xin have you reviewed the PR?

[xin-zhang2]

The surviving rows are already reflected in numRows, so there is no need to check parentNulls. We can simply use

        while (numRows > 0) {
          auto n = std::min<vector_size_t>(numRowsPerChunk, numRows);
          outputStreams[p]->write(chunkBytes, n * sizeof(T));
          numRows -= n;
        }

regareless of whether parentNulls is null or not.
Then parentNulls can be removed from the function parameters.

[xin-zhang2]

for (uint32_t p : nonEmptyPartitions)

[xin-zhang2]

outputStreams[p] != nullptr is unnecessary here.

[xin-zhang2]

It looks hasParentNulls is redudant.
parentNulls is always resized in the code, so parentNulls[vectorIndex] == nullptr is enough to check whether the parent has nulls.

[xin-zhang2]

We can add a few fast paths here. The following change shows a slight improvement in the benchmark for with-null cases.

const uint64_t numBits = end - begin;
  if (mask == nullptr) {
    if (source == nullptr) {
      bits::fillBits(target, targetBitOffset, targetBitOffset + numBits, bits::kNotNull);
    } else {
      bits::copyBits(source, begin, target, targetBitOffset, numBits);
    }
    return static_cast<int32_t>(numBits);
  }

  if (source == nullptr) {
    const int32_t count = bits::countBits(mask, begin, end);
    bits::fillBits(target, targetBitOffset, targetBitOffset + count, bits::kNotNull);
    return count;
  }

  uint64_t outBit = targetBitOffset;
  bits::forEachWord(begin, end, [&](int32_t index, uint64_t wordMask) {
    const uint64_t selected = mask[index] & wordMask;
    if (selected == 0) {
      return;
    }
    const uint64_t packed =
        bits::extractBits<uint64_t>(source ? source[index] : ~0ULL, selected);
    const uint32_t count = __builtin_popcountll(selected);
    appendLowBits(target, outBit, packed, count);
    outBit += count;
  });
  return static_cast<int32_t>(outBit - targetBitOffset);

[xin-zhang2]

We may need to consider the encoding before casting. A ROW type vector can be encoded as CONSTANT, in which case pv would be a PartitionedConstantVector instead of PartitionedRowVector.
The same issue also applies at line 210.

[xin-zhang2]

I wonder if it make sense to split the footer writing into separate functions? Since different encodings may share the same wire-format pieces, such as offsets and null bitmaps, this logic might be reusable.