[C++][Parquet] Reuse memory of passed array in ConvertToDecimal

[Hattonuri]

Describe the enhancement requested

I am reading an arrow::ListArray of Decimal128 from FixedLengthBinaryArray[16] in parquets and see on flamegraph that RawBytesToDecimalBytes consumes large amount of time and mostly due to page faults.

I assume that the problem happens here

arrow/cpp/src/parquet/arrow/reader_internal.cc

Lines 559 to 576 in eb5de18

	// allocate memory for the decimal array
	ARROW_ASSIGN_OR_RAISE(auto data, ::arrow::AllocateBuffer(length * type_length, pool));
	// raw bytes that we can write to
	uint8_t* out_ptr = data->mutable_data();
	// convert each FixedSizeBinary value to valid decimal bytes
	const int64_t null_count = fixed_size_binary_array.null_count();
	using DecimalType = typename DecimalTypeTrait<DecimalArrayType>::value;
	if (null_count > 0) {
	for (int64_t i = 0; i < length; ++i, out_ptr += type_length) {
	if (!fixed_size_binary_array.IsNull(i)) {
	RETURN_NOT_OK(RawBytesToDecimalBytes<DecimalType>(
	fixed_size_binary_array.GetValue(i), byte_width, out_ptr));
	} else {
	std::memset(out_ptr, 0, type_length);
	}

when arrow allocates new buffer for decimals and then parses FLBA into it.
As FLBA type has the same amount of memory per element i guess that this function should reuse given array data.

Possible easy(but very dirty) solution could be static_cast given array to Decimal array (as it does not have additional fields - it wouldn't be an error), const_cast raw_values of it and reverse byte order in-place
Another benefit is reducing memory usage of reading them

Component(s)

C++, Parquet