issue #619: bulk-prefetch vector segment warmup via SegmentBlockCache.bulkInsert

[eolivelli]

Fixes #619.

What

PersistentVectorStore.warmUpNewSegmentsBeforePublish is the dominant cost of
checkpoint Phase C-prep and compaction publish on the indexing service —
because the warmup BFS issues one ~16 KiB readFileRange RPC per HNSW
Layer-0 node, ~2000 round-trips per 32 MiB-budgeted segment, against the
remote file server / S3 / GCS.

This PR replaces those many small reads with a small number of large
multipart-chunk-sized bulk reads that go straight into the
SegmentBlockCache. The BFS still runs afterwards — exclusively against the
hot cache — so it preserves the per-segment warmedUp accounting and the
pin BFS pass for the frontier region (#578), with no behaviour change for
local-file / in-memory readers.

Changes

• BulkPrefetchReaderSupplier (new, in herddb-core): mixin interface
  analogous to PinModeReaderSupplier. Exposes
  long bulkPrefetchIntoCache(long startOffset, long maxBytes).
• SegmentBlockCache.bulkInsert(path, baseOffset, blockSize, ByteBuf):
  new API that splits a contiguous prefetched range into block-sized retained
  slices keyed identically to getBlock, taking ownership of the input buffer
  and never overwriting a single-flighted entry. New stats counters
  bulkInsertedBlockCount / bulkInsertedByteCount / bulkInsertSkippedCount.
• RemoteRandomAccessReader.Supplier now implements
  BulkPrefetchReaderSupplier: dispatches one
  readFileRangeAsByteBufAsync per multipart chunk in parallel, then
  bulk-inserts each completed chunk into the cache. Aligns to the
  writeBlockSize boundary because the storage backend forbids reads
  spanning more than one multipart chunk.
• PersistentVectorStore.bulkPrefetchSegmentIntoCache: invoked from
  warmUpSegment before the BFS. Best-effort: an IOException from the
  supplier is logged at WARNING and the original per-node BFS path still
  runs, so we never regress versus pre-Speed up segment warm up #619 behaviour even when the remote
  storage is degraded.

Tests

• SegmentBlockCacheBulkInsertTest (10 tests): contract of the new
  bulkInsert API — full-block range, partial last block, argument
  validation (always releases the input buffer), disabled-cache no-op,
  existing-entry preservation + skip counting, invalidatePath drives the
  parent buffer's refCnt back to 0 (no leak), non-zero aligned offsets,
  repeated insert.
• RemoteRandomAccessReaderBulkPrefetchTest (11 tests): end-to-end
  against an in-process RemoteFileServer + multipart file spanning 3 chunks
  — supplier implements the mixin; full prefetch leaves the cache hot with
  zero loader invocations; partial prefetch only caches the requested chunks;
  disabled cache short-circuits; missing path surfaces as IOException;
  repeated prefetch counts skips; invalidatePath releases every
  prefetched block.
• Issue619BulkWarmupTest (7 tests): dispatch contract from
  PersistentVectorStore.bulkPrefetchSegmentIntoCache — exactly-once call
  with (0, budget) for matching suppliers, no call for plain ReaderSupplier,
  IOException and zero/negative budget swallowed, segment state not
  mutated.

Test plan

• mvn -pl herddb-remote-file-service -Dtest='SegmentBlockCacheBulkInsertTest,RemoteRandomAccessReaderBulkPrefetchTest,SegmentBlockCacheTest,SegmentBlockCacheAsyncTest,SegmentBlockCacheHammerTest' test — 57/57 ✅
• mvn -pl herddb-indexing-service -Dtest='Issue619BulkWarmupTest,Issue569WarmupBeforePublishTest,BlockCacheWarmupTest' test — 19/19 ✅
• LazyValueCacheConcurrentUpdatesSuite{NoIndexes,WithNonUniqueIndexes,WithUniqueIndexes}Test — 12/12 ✅
• mvn -B spotless:check apache-rat:check spotbugs:check install -DskipTests -Pci — BUILD SUCCESS ✅
• CI

🤖 Implemented by the pr-worker agent.