[core] Release every reader in SortMergeReaderWithMinHeap.close()

[PDGGK]

Purpose

Closes #9162.

SortMergeReaderWithMinHeap.close() releases one RecordReader per sorted run being merged, across three loops, and every call was a bare close(). The first one to throw abandoned the rest of its loop and both loops after it.

This is not a two-resource teardown: there is one reader per sorted run, each holds an open data file, and it runs on the compaction path. A single unreadable file therefore strands every descriptor behind it for the remainder of the merge. element.iterator.releaseBatch() had the same exposure — no checked exception, but it can still throw, and when it did the close() on the next line was skipped.

Each step now runs, the first failure propagates, and later ones are attached to it as suppressed.

On the choice of helper. IOUtils.closeAll is the obvious candidate and I did not use it: it declares throws Exception, and both existing callers (FormatTableFileWriter:93, FileChannelManagerImpl:126) widened their own signature to match. This method overrides RecordReader.close() throws IOException and cannot widen. ExceptionUtils.firstOrSuppressed is the right fit — its javadoc gives this exact close-loop as its worked example.

SortMergeReaderWithLoserTree needs no change; it delegates to a single loserTree.close().

Tests

New SortMergeReaderWithMinHeapCloseTest, three cases:

case	asserts
an earlier reader fails to close	all three readers still closed; the first failure propagates with the later one as suppressed
releaseBatch() throws	both readers still closed, and the release failure is what the caller sees
everything closes cleanly	no exception — the control

Against the previous close(), the first and third of those fail; the all-succeed control passes on both versions, which is what shows the other two aren't trivially red.

The readers are a small hand-rolled fake rather than mocks, so the test asserts on real call ordering through readBatch() into the heap rather than on stub interactions.

spotless:check, checkstyle:check and apache-rat:check are clean.

Local run caveat. With -Pfast-build, CompactionChangelogFollowUpScannerTest and MergeTreeCompactManagerTest error in setup with Could not instantiate generated class 'RecordComparator' — the codegen module isn't built under that profile. I confirmed it's unrelated by reverting SortMergeReaderWithMinHeap.java to master and deleting the new test: both fail identically. Everything else in the merge-tree and compaction packages passes, including MergeTreeReadersConnectionsLeakTest (20).