Engine capability registry for function availability

[shangyian]

Currently, engines can be registered with freeform names and versions, but DJ has no way to know what functions are available on each engine. This leads to potential issues:

• Decomposition may generate unsupported SQL - e.g., using hll_sketch_agg on Spark 3.3 which doesn't have it
• No graceful degradation - DJ can't fall back to simpler approaches when advanced features aren't available
• Function translation is fragile

Examples

Function	Spark 3.3	Spark 4.0+	Druid
hll_sketch_agg	✅	✅	✅ (DS_HLL)
theta_sketch_agg	❌	✅	✅ (THETA_SKETCH)

Proposed Solution

Replace freeform engine registration with a curated list of supported engine/dialect combinations:

SUPPORTED_ENGINES = {
    "spark:3.5": SparkDialect35(),
    "spark:4.0": SparkDialect40(),  
    "trino:4xx": TrinoDialect(),
    "druid:31": DruidDialect(),
}

Each dialect would declare:

• Available functions
• Function name mappings for translation
• Valid decomposition strategies
• Type coercions