Support `APPROX_COUNT_DISTINCT ` with HLL sketches for cross-engine approximate count distinct

[shangyian]

Introduce APPROX_COUNT_DISTINCT(...) as a first-class aggregation type that uses HyperLogLog sketches under the hood. HLL sketches are composable - they can be merged to produce combined cardinality estimates with ~1-2% error.

name: unique_customers
type: metric
expression: APPROX_DISTINCT(customer_id)

Materialization (Build Time)

DJ generates sketch accumulation SQL:

-- Spark
SELECT date, region, hll_sketch_agg(customer_id, 12) as cust_sketch
FROM orders GROUP BY date, region

Query Time (Rollup)

When querying across materialized data, DJ generates merge + estimate:

-- Druid
SELECT
  HLL_SKETCH_ESTIMATE(HLL_SKETCH_UNION(cust_sketch))
FROM materialized_cube

Fallback (No Materialization

SELECT hll_sketch_agg(customer_id, 12) FROM orders