misc: Sync upstream Velox main

[firestarman]

Merge summary

• Upstream range: 796c6214c89a78ef86df29a64d4351129f62cb86..d13ef1e303af6a12cfb770f8144fad49594bd5ba
• Original sync base: open PR head f95611c3ee9d3556e2a600aaf1faef20d487aba9
• Recovered unpublished range: none
• Upstream integration: merge commit 2cb8ace70220dc8a151816ab1412ea472b6a1202
• Latest downstream dev: freshly fetched cf619d90eca1b83ad31f446b96482df096d24923 was already contained; no second merge commit was required.
• Final Velox commit: 2cb8ace70220dc8a151816ab1412ea472b6a1202
• .github: retained exactly from the recorded latest downstream/dev; upstream .github changes were ignored.

Notable changes

• Torchwave adds per-column decomposition and chain-gather folding, a write/read barrier fix, fused sparse flip-and-truncate support, a standalone Triton masked_select_jagged, and .pt2 flat-input loading.
• Nimble vectorizes ALP encoding and grid search, hardens FSST decode bounds and symbol isolation, synchronizes encoding-pool callbacks, skips index-projection stripes with no projected streams, structurally slices partial BlockBitPacking blocks, and exposes writer options through SerDe keys.
• Spark SQL adds AverageAggregate::toIntermediate and btrim; vector decoding exposes whether buffers are borrowed or owned.
• Connector downcasts and table-scan plan text were improved, and deprecated gMock Invoke wrappers were removed. This range has no net experimental/cudf changes.
• No downstream Velox compatibility commit was needed after the merge. NVIDIA/spark-gluten#172 remains required to adapt Spark-Gluten's plan converter to the writer and Unnest APIs present at this Velox head.

Conflict resolutions

• No manual Git conflicts.
• .github was restored and verified against the recorded baseline after the upstream merge, per downstream-ownership policy.

Validation

• Status: BLOCKED_ON_SPARK_GLUTEN_FIX — smoke and full Colt validation pass against the unmerged fix below; this Velox PR remains Draft until that fix lands in Spark-Gluten dev and validation is rerun from scratch against dev.
• Spark-Gluten dev: 9f732196834cb3d1f67488dd1d2f93823802dc7a
• Spark-Gluten fix: NVIDIA/spark-gluten#172, tested SHA df462d247c8f5193d578ee8f8242d1a8c909748e (Ready).
• Colt: 73f99f12c9f0fe067f48a667a68e0c52e4dbd832
• Verification config: ebb16dcd6fcb238b34d39cd00a2eb46a9525d5ec; generated manifest 9b45526e6ac686c3fb1a95fdb2f94f1ae48e90b7113de4dcfe80d58f5835e338; provenance at /home/liangcail/velox-upstream-sync-20260814/results/20260907T041509Z/verify-config/provenance.json (SHA-256 e28bcbb0f83e7c716050f5468e377e3ddc8878276d763ca0b78ba5f139a0732d).
• Initial current-dev failure: build-libs, return code 2. InsertTableHandle required notNullColumns, and UnnestNode required optional output names. Evidence: /home/liangcail/velox-upstream-sync-20260814/results/20260907T033307Z/smoke/verify/workflow/logs/10-build-libs/spark-gluten-build-libs.log.
• Repair: Spark-Gluten df462d247c8f5193d578ee8f8242d1a8c909748e supplies an empty notNullColumns set to preserve existing behavior and uses vector<optional<string>> for Unnest names. git diff --check, both independent clean consumer builds, and all downstream tests passed.
• Smoke, TPC-H Query 1: PASS, 361 source/native cases recorded (235 passed, 126 suite-declared skips) and Query 1 passed; /home/liangcail/velox-upstream-sync-20260814/results/20260907T041509Z/smoke/verify.
• Full, TPC-H Query 1–22: PASS, 361 source/native cases recorded (235 passed, 126 suite-declared skips) and all 22 queries passed; /home/liangcail/velox-upstream-sync-20260814/results/20260907T041509Z/full/verify.
• Validation receipt: /home/liangcail/velox-upstream-sync-20260814/results/20260907T041509Z/sync-validation.json.
• CUDA/GPU/images: CUDA 12.9; requested native, detected SM89 on GPU 0 (NVIDIA L4); driver mode auto resolved to host; cuDF 5beaa5954688fcb12236ffb434e192ea2c77db30; sealed dependency carrier sha256:b52df2b7521e0fb00e0ec14a3cf7428c7bbd8dba0caba037c5d308c9c19b3094; smoke/full build execution images sha256:b13917dd0d88e30d94aab058e18e754e419348c3f679c26cf05f8583f71a3782 and sha256:e2724244a0268243dfc11747b45c29d6c45ddfac5f5f30ea99769328e3c8828a; smoke/full runtime image sha256:1b42a229c8357223f5c3bdc06dceee2ddddf6c3d0e692db1b5a8b8b074ffaaa3.