Avoid associative (and identity) aggregations over the complete partition

[lukaseder]

Associative and identity aggregations could be calculated based on the result of the aggregation of the previous value:

v   sum (-3, -1)  current calculation   better
-------------------------------------------------
1   e             e                     e
2   1             1                     e + 1
3   3             1 + 2                 1 + 2
4   6             1 + 2 + 3             3 + 3
5   9                 2 + 3 + 4         6 + 4 - 1
6   12                    3 + 4 + 5     9 + 5 - 2

As the partition is already sorted correctly, we can simply access the (cached) calculated value from lag()

Associative and identity operations currently supported are:

• count() (optimisation not necessary, as count() is already O(1))
• sum() (and all derived aggregations)
• avg() (and all derived aggregations - not strictly associative, but can depend on sum() and count())
• toList() (a potential optimisation would require an immutable linked list)
• toString()

More limited optimisation is possible for associative-only aggregations, when using a Long.MIN_VALUE .. 0 frame. These include:

• min()
• max()
• median() (would depend on cached toList())
• percentile() (would depend on cached toList())
• allMatch()
• anyMatch()
• noneMatch()