Combine accumulator and scaler to avoid communication across execution nodes

[mIXs222]

Previously, each scaler is a separate execution node subscribing to the previous aggregation accumulator and subscribed by the next node. This could lead to unnecessary cloning and communication across nodes, as well as overly complicated query implementations.

• Pull out the common trait on common scalers and count distinct scaler
• Add a scaler as an attachment to the accumulator