Avoid retaining intermediate SQL reporting rows

[gregjotau]

The SQL backend currently loads every row into a cloned DBResultRow collection, then builds the PeppolReportingItem collection in a second pass. Large reporting periods retain both the intermediate row/column wrappers and the final objects.

Use the existing DBExecutor.queryAll row callback to construct each reporting item immediately. The SQL, date bounds, column conversions and eager Iterable API remain the same; no dependency, JDBC fetch setting, cursor API or schema changes are needed.

The query's ESuccess is checked before returning the collection. A failed query now throws the already-declared PeppolReportingBackendException, rather than returning an empty report as before; failure after some rows also cannot expose a partial collection. The database executor continues to log the underlying error.

Measured against main 55153b3 using local PostgreSQL 18.6 and 200,000 synthetic rows (the existing PeppolReportingTestItems.sendingItem with distinct end-user IDs), OpenJDK 26.0.2.1/macOS arm64, -Xms64m -Xmx768m, five alternating fresh JVM runs per version:

Median	Before	After
Peak process RSS (/usr/bin/time -l)	642.2 MiB	496.1 MiB
Query + item mapping elapsed time	387.6 ms	342.0 ms
Current-thread query allocation	443.6 MiB	423.7 MiB
Heap after GC with the returned items retained	165.3 MiB	165.4 MiB

All runs returned 200,000 rows with identical checksums. This local workload uses about 146 MiB (23%) less peak process memory. The final collection size is unchanged; this removes intermediate retention, not the eager collection or JDBC driver's own buffering.

Validation:

• mvn -B -pl peppol-reporting-backend-sql -am -Dtest=SPITest,PeppolReportingBackendSqlTest -Dsurefire.failIfNoSpecifiedTests=false clean verify: all four reactor modules pass, three selected tests, no failures/errors/skips.
• All 16 existing shared backend contract tests run against the changed SQL backend on an isolated local PostgreSQL database, with zero failures or skipped assumptions. These cover field round trips, inclusive dates, multiple rows, empty results, callbacks and lifecycle behavior. Other database dialects were not exercised.
• The new database-independent regression test verifies query failure cannot be mistaken for an empty report. It fails against the original implementation and passes with this change.

[phax]

Thanks, another quick win :)