test: move explode/posexplode coverage from CometGenerateExecSuite to Comet SQL tests #6630

Part of #6615.

What / Why

This issue moves the generator coverage in exec/CometGenerateExecSuite to Comet SQL tests: explode, explode_outer, posexplode, posexplode_outer, over arrays and struct arrays, across batch boundaries, and under LIMIT ... OFFSET. Fixtures go under sql-tests/expressions/array/, next to explode.sql and posexplode.sql. Static analysis on main at 3bc2faa found the following:

Nearly everything converts. The plan-shape test and the ScalaUDF test stay.

Work

Line numbers refer to CometGenerateExecSuite.

• explode_local_table_scan.sql (new), covering the 24 Seq.toDF tests from L114 to L436 except L214, L227 and L393
  • Use -- Config: spark.comet.exec.localTableScan.enabled=true with inline VALUES.
  • The behavior is already in explode.sql / posexplode.sql over Parquet. What this file adds is the source: Parquet always reads arrays back with containsNull=true, while Seq.toDF gives non-null elements, which is the bug: array/map kernels crash on non-null child field via CometLocalTableScanExec #4789 hazard.
  • ConvertToLocalRelation folds only Project, Filter and Limit, so a Generate over VALUES still reaches the native explode.
  • Write NULL arrays as bare NULL, because CAST(NULL AS ARRAY<INT>) makes the elements nullable again.
  • Seven of these behaviors are in no fixture today: L180, L191, L202, L240, L297, L381 and L436.
• explode_disabled.sql (new, L214): set spark.comet.exec.explode.enabled=false and assert the fallback reason.
• explode.sql (L227): add the explode(m) fallback.
  • Also move the rand() query out into its own file. The whole file is gated MinSparkVersion: 3.5 for that one query, so today it is skipped entirely on Spark 3.4.
• posexplode.sql (L100): add posexplode(split(s, 'x')).
• explode_small_batches.sql (new): L447, L465, L488, L613, L629, L646, using -- ConfigMatrix: spark.comet.batchSize=4,8 and single-file tables.
• explode_limit_offset.sql (new): L523, L539, L561, L580, using single-file tables with LIMIT 4 OFFSET 1 (ListPositionsExpr panics on sliced list input, breaking native posexplode over LIMIT with OFFSET #5224).
• explode_boolean_struct.sql (new): L691 (four instances) and L710, using a 3000-row table generated with hash() (Native explode returns wrong booleans for arrays of structs after the first batch of output #6464).
• explode_boolean_struct_batch_100.sql (new): L700, with spark.comet.batchSize=100.

Use INSERT INTO t SELECT /*+ COALESCE(1) */ ... or range(0, n, 1, 1) for every batch-boundary fixture. INSERT ... VALUES writes one file per row and would never put two rows in one batch.

Stays in Scala

• L38 (8 instances) pins CometExplodeExec canonicalization (outer and non-outer must differ, identical generators must match) and a ReusedExchangeExec over the generator with AQE on. That is plan shape.
• L717 feeds an exploded boolean into a Scala UDF (Native explode returns wrong booleans for arrays of structs after the first batch of output #6464), and UDF registration has no SQL form. The rest of that shape is covered by L710.

Notes on the Scala tests

• The non-outer NULL/empty tests never reach the native operator with those rows (L125, L136, L263, L323, L334, L406). Spark's InferFiltersFromGenerate adds a non-NULL, non-empty filter, and ConvertToLocalRelation applies it to the Seq data at plan time. The fixtures keep the queries, but the coverage that matters is in the _outer variants.
• L447 never has two input batches in one partition.
• Bare -0.0. explode.sql:142,156 and posexplode.sql:131 write cast(-0.0 as float/double), which is +0.0. Use float('-0.0') / double('-0.0').

Done when

• Every test in the table below is accounted for. Either a fixture covers it (the same inputs, configs and assertion strength, or stronger), the fixture that already covers it is cited, or it stays in Scala for the reason given above.
• Each converted Scala test is deleted in the PR that adds its coverage. A suite left empty is deleted and removed from the suite lists in .github/workflows/pr_build_linux.yml and pr_build_macos.yml.
• The new and changed fixtures pass on every Spark profile. Apply run-all-spark-profiles, because pull requests run only the default profile.