fix(e6): preserve DBR DATE_ADD arity under E6_EXECUTOR_TYPE=native

[tkaunlaky-e6]

Summary

• Databricks 2-arg DATE_ADD(date, n) returns DATE; 3-arg DATE_ADD(unit, n, ts) returns TIMESTAMP. The transpiler previously collapsed both to e6 3-arg DATE_ADD('DAY', n, ts), switching the return type and breaking CONCAT(DATE_ADD(ts, 0), SUBSTR(ts, 11)) patterns by producing a doubled time portion (Samsung zero-row issue).
• New behavior is gated behind E6_EXECUTOR_TYPE=native (the same flag used by PR fix(e6): gate TO_UNIX_TIMESTAMP /1000 on executor type #253). Java executor default is unchanged.
  • E6_EXECUTOR_TYPE=native: 2-arg DBR -> e6 2-arg DATE_ADD(ts, n); 3-arg DBR -> e6 3-arg DATE_ADD(unit, n, ts).
  • default (java): 2-arg and 3-arg DBR both emit e6 3-arg with DAY filled in (prior behavior preserved).
• Non-DAY units (e.g. DATE_ADD('YEAR', 5, d)) always stay 3-arg in both modes.

Mechanism

• sqlglot/dialects/databricks.py: pass default_unit=None to build_date_delta for DATE_ADD/DATEADD. Now 2-arg DBR parses with unit=None, 3-arg with an explicit unit. Matches the pattern at clickhouse.py:310. No DBR output change since date_delta_sql falls back to DAY via unit_to_var.
• sqlglot/dialects/e6.py: new auto-discovered dateadd_sql method reads E6_EXECUTOR_TYPE and emits 2-arg only when native AND unit is None. Replaces the previous exp.DateAdd: lambda TRANSFORMS entry. Read parser also handles 2-arg DATE_ADD(date, n) for round-trips.

Context

The Samsung repro pattern:

CONCAT(
  DATE_ADD(CAST(... AS TIMESTAMP), CAST(7 * 0 AS INT)),
  SUBSTR(<ts>, 11)
)

With prior 3-arg conversion, DATE_ADD returned TIMESTAMP -> CONCAT(<full ts>, ' 00:00:00') -> doubled time portion -> invalid timestamp -> NULL after outer CAST, predicate becomes unknown, rows dropped.

Test plan

• New test_date_add_native_executor covering both modes (2-arg/3-arg input, 2-arg/3-arg expected output).
• Existing test_dialect_specific_functions extended with the java-default case (DATE_ADD(ts, 2) -> DATE_ADD('DAY', 2, ts)).
• python -m unittest discover tests/dialects -> 545/545 pass (e6, databricks, spark, hive, etc.) -- no regressions.