Ordering by grouping column performs unnecessary work

[cswinter]

Note this portion of the query plan which combines passenger_count with itself:

casted_1     = column_0 as I64                               TypeConversionOperator<u8, i64>
bitpacked_2  = casted_1 + (casted_1 << $shift)               ParameterizedVecVecIntegerOperator<BitShiftLeftAdd>

This causes a roughly 3x slowdown (on this query). When creating the grouping key, each unique expression should just be used once.

Full query plan:

locustdb> :explain SELECT passenger_count, count(0) FROM trips ORDER BY passenger_count LIMIT 100;

Query plan in 22412 batches
-- Stage 0 (streaming) --
column_0     = "passenger_count".0                           ReadColumnData
casted_1     = column_0 as I64                               TypeConversionOperator<u8, i64>
bitpacked_2  = casted_1 + (casted_1 << $shift)               ParameterizedVecVecIntegerOperator<BitShiftLeftAdd>
constant_3   = Constant<Integer>                             Constant
count_4[bitpacked_2] += 1                                    VecCount<i64>

-- Stage 1 --
nonzero_indices_5 = nonzero_indices(count_4)                 NonzeroIndices<u32, i64>

-- Stage 2 --
             = count_4[count_4 > 0]                          NonzeroCompact<u32>

-- Stage 3 --
casted_6     = count_4 as I64                                TypeConversionOperator<u32, i64>

-- Stage 4 --
unpacked_7   = (nonzero_indices_5 >> $shift) & $mask         BitUnpackOperator
casted_8     = unpacked_7 as U8                              TypeConversionOperator<i64, u8>
casted_9     = casted_8 as I64                               TypeConversionOperator<u8, i64>

-- Stage 5 --
unpacked_10  = (nonzero_indices_5 >> $shift) & $mask         BitUnpackOperator
casted_11    = unpacked_10 as U8                             TypeConversionOperator<i64, u8>
casted_12    = casted_11 as I64                              TypeConversionOperator<u8, i64>